AROBUSTMETHODFORADDRESSINGPUPILDILATIONINIRISRECOGNITIONByRaghunandanPasulaATHESISSubmittedtoMichiganStateUniversityinpartialoftherequirementsforthedegreeofComputerScienceŒMasterofScience2016ABSTRACTAROBUSTMETHODFORADDRESSINGPUPILDILATIONINIRISRECOGNITIONByRaghunandanPasulaTherichtextureoftheirisisbeingusedasabiometriccueinseveralhumanrecognitionsystems.
Irisrecognitionsystemsarefairlyrobusttosmallchangesinilluminationandpose.However
thereareanumberoffactorsthatstilladverselyaffecttheperformanceofanirismatcher.These
includeocclusion,largedeviationingaze,lowimageresolution,longacquisitiondistanceand
pupildilation.Largedifferencesinpupilsizeincreasesthedissimilaritybetweenirisimagesofthesameeye.Inthiswork,thedegradationofmatchscoresduetopupildilationissystematicallystudiedusingHammingDistancehistograms.Anovelrule-basedfusiontechniquebasedontheaforemen-
tionedstudyisproposedtoalleviatetheeffectofpupildilation.Theproposedmethodcomputesa
newdistancescoreateverypixellocationbasedonthesimilaritiesbetweenIrisCodebitsthatwere
generatedusingGaborFiltersatdifferentresolutions.Experimentsshowthattheproposedmethodincreasesthegenuineacceptratefrom76%to90%at0:0001%falseacceptratewhencomparingimageswithlargedifferencesinpupilsizesintheWVU-PLRdataset.Theproposedmethodisalsoshowntoimprovetheperformanceof
irisrecognitiononothernon-idealirisdatasets.Insummary,theuseofmulti-resolutionGabor
Filtersinconjunctionwitharule-basedintegrationofdecisionsatthepixel(bit)levelisobserved
toimprovetheresilienceofirisrecognitiontodifferencesinpupilsize.ACKNOWLEDGEMENTSIwouldliketothankDr.ArunRossforhiscontinuedsupportandguidancethroughoutmystudent
career.Noamountofthankswouldbesuftodescribethesupportextendedbythefamily
membersbyjustbeingthereformeandsupportingmethroughthedegreeprogram.SpecialthankstoDr.EricTorngandKatherineTrinkleinforhelpingmethroughthetoughtimes.Thisworkisdedicatedtoalltheguruswhospendamountofeffortandtimetoimpartknowledgetotheworld.iiiTABLEOFCONTENTSLISTOFTABLES.......................................viLISTOFFIGURES.......................................viiCHAPTER1INTRODUCTION...............................11.1Biometrics.......................................1
1.2Eyeanatomy......................................41.2.1Layersofeye.................................5
1.2.2Apparentiristexture.............................9
1.2.3Pupildynamics................................121.3Irisbiometricsystem.................................13
1.4ChallengesinIrisrecognition.............................20
1.5Objectivesofthiswork................................25CHAPTER2MOTIVATIONANDPREVIOUSWORK..................262.1Motivation.......................................26
2.2Previouswork.....................................282.2.1Minimumwearandtearmodel........................282.2.1.1Wyatt2000.............................28
2.2.1.2YuanandShi2005.........................31
2.2.1.3Weietal.2007...........................322.2.23-Danatomicalmodel............................332.2.2.1Francoisetal.2007........................33
2.2.2.2Clarketal.2012..........................342.2.3Gejjietal.2015................................36
2.2.4OtherDeformationmodels..........................362.3Bitmatching......................................37CHAPTER3COLLECTIONOFDATABASE........................393.1Motivation.......................................39
3.2Dataacquisitionprotocol...............................39
3.3Description......................................40
3.4Impactofpupildilation................................42CHAPTER4PROPOSEDMETHODS...........................444.1Multi-resolutionGaborencoding........................44
4.2TypicalIrisCodematcher...............................45
4.3Histogramofmatchingpatterns............................46
4.4Fusion.........................................484.4.1RulebasedFusion...............................49
4.4.2basedFusion............................524.5ExperimentsandResults...............................52iv4.6Examples.......................................56CHAPTER5SUMMARY..................................605.1Summary.......................................60BIBLIOGRAPHY........................................61vLISTOFTABLESTable1.1Wavelengthrangeforvisible,NIRandSWIRspectrum..............9
Table3.1Demographicsdistribution.............................41
Table3.2Eyecolorinformation...............................41
Table4.1LogicaloperationsusedtocombinetheoutputofmultipleIrisCodes.......52viLISTOFFIGURESFigure1.1FigureshowingtheexternalanatomyofthehumaneyeintheRGBspectrum.Thefocusofthisthesisisontheiris,whichistheannulartexturedstructure
situatedbetweenthepupilandthesclera.Theirisistypicallyimagedinthe
nearinfrared(NIR)spectrumandnotintheRGBspectrum...........2Figure1.2Abiometricsystem,duringthe,enrollmentstageaddsatemplatebelongingtoanewuserintotheGallery............................3Figure1.3Sagittarialcross-sectionofiris.Imagepublishedherewithpermissionfrom[1]6
Figure1.4Differentlayersofiriswhenlookingintosagittalaxis...............7
Figure1.5Locationofsphincteranddilatormusclesthatcontrolpupilconstrictionanddilation,respectively................................8Figure1.6Pathfromlightsourcetotheeye.........................10
Figure1.7Figureshowingwavelengthsofelectro-magneticspectrumrelevanttoirisbiometrics.....................................11Figure1.8Absorptionspectrumof(a)liquidwater[2]and(b)melanin[3]atdifferentwavelengthsofelectromagneticspectrum....................11Figure1.9Exampleofblue,yellowishanddarkbrownirisimagesthatcontainlow,moderateandhighconcentrationofmelanin,respectively............12Figure1.10Darkirisin(a)imagedat(b)470nm,(c)520nm,and(d)700nmand(e)NIRwavelengths....................................12Figure1.11Examplesoffactorsthesizeofthepupil...............13
Figure1.12Componentsofatypicalirisrecognitionsystem.................14
Figure1.13(a)Realpartand(b)ImaginarypartofaGabor...............18
Figure1.14(a)OriginalacquiredImage(b)Segmentationoutput(c)Normalizedimage(d)Correspondingmaskimage(e)IrisCodegeneratedbyencodingthenor-
malizedimageusingMasek'smethod.[4]....................19Figure1.15Examplesofnonidealirisimages.(a)and(b)Non-uniformillumination,(c)and(d)Eyelidocclusion,(e)Eyelashocclusion,(f)Motionblur.........22Figure1.16Examplesofoff-axisirisimages..........................22viiFigure1.17Fewexamplesofeyediseasesthatimpactirisrecognition(a)Polycoria-Multiplepupilopenings(b)Coloboma-Teariniris(c)Severecataract-
Thickeningoflens(loosestransparency).Althoughtheimagesareshownin
RGB,someofthesediseasescanalsoimpacttheNIRimages...........24Figure2.1(a)and(b)Irisimagewithmoderatepupilsizeandthecorrespondingnor-malizedirisimage.(c)and(d)Irisimagewithlargepupilsizeandthecor-
respondingnormalizedirisimage.Highlightedregionsin(c)and(d)donot
aligncorrectly.Imagesfrom[5]..........................27Figure2.2Irisimageswithdilationratiosof(a)0:3478and(b)0:6545.Imagesfrom[6]..27Figure2.3IrismeshworkproposedbyRohen[7].Imagefrom[7]..............29
Figure2.4qoistheanglebetweenstartingpointofthearconpupillaryboundaryandendingpointonlimbicboundary.......................30Figure2.5OptimumarcsderivedbyWyatt[8]forq=100,andpupildiameter1.5,4.0and7.0mm....................................30Figure2.6NormalizationmodelproposedbyYuanandShi.Imagefrom[9]........31
Figure3.1Imagesequencecapturestartsatt0=0.Afterapproximately10seconds,att1,thelightsourceisturnedonilluminatingtheeyefor10moreseconds[t1;t2].Att2thelightsourceisturnedoffandremainsofffor10moreseconds[t2;t3].Thevideocaptureisstoppedatt3......................40Figure3.2Sampleimagesfromthedataset..........................41
Figure3.3Distributionofpupildilationratiosinthedataset.Theyrangefrom0.2177to0.6367........................................42Figure3.4DistributionofgenuineHammingdistancescoresasafunctionofdilationdifferences.(a)jD1D2jand(b)jR1R2j...................43Figure4.1Anormalizedimageisencodedusingmulti-scaletoresultinanIrisCodesetalongwithamaskshowingvalidbitsineachIrisCode.Thismaskissame
forallthecodesintheIrisCodeset........................45Figure4.2AnormalizedimageanditscorrespondingIrisCodegeneratedusing3Theseencodetheimageatmultiplescales.................45Figure4.3Atypicalirismatcher.Matchscoresarecomputedindependentlyateachscalewhicharethenfusedatscoreleveltoresultinadistancescore....47viiiFigure4.4Distributionofdecisionsforgenuinematchingcasesforasinglesubject........................................48Figure4.5Distributionofdecisionsforrandomlyselectedimpostormatch-ingcases......................................49Figure4.6Comparisonofdistributionsofpossibledecisionsforgenuineandimpostormatchingcases.............................50Figure4.7Theproposedirismatchersequentiallycombinestheresultsatmultiplescalesandgeneratesasingledecisionresult........................51Figure4.8FlowchartdepictingMethod1anditscorrespondingtruthtable.........53
Figure4.9FlowchartdepictingMethod2anditscorrespondingtruthtable.........54
Figure4.10FlowchartdepictingMethod3anditscorrespondingtruthtable.........55
Figure4.11(a)ROCsforfulldata.Thegenuineandimpostorscoredistributionsareplottedfor(b)Method1,(c)Method2and(d)Method3.............56Figure4.12ROCsgeneratedbyusingthegenuinescoresforpairswhosepupildilationratiodifferencesare(a)small,(b)mediumand(c)large.Theimpostordis-
tributionsareheldthesameacrossallthecases..................57Figure4.13ThehistogramofgenuineandimpostorscoresusingMasek'smethodandafterfusionofmatchscoresfromMasek'smethodandproposedMethod1....58Figure4.14ROCcurvesfor(a)WVUand(b)QFiredatasets.TheimprovementinGARisclearlyevidentatlowFARs...........................58Figure4.15Genuinepairsofimagesthatwerecorrectlymatchedusingtheproposedmethodbutwereincorrectlyrejectedbythetraditionalmatchingmethodat
0:0001%FAR...................................59ixCHAPTER1INTRODUCTIONfiTosupposethattheeyewithallitsinimitablecontrivancesforadjustingthefocus
todifferentdistances,foradmittingdifferentamountsoflight,andforthecorrection
ofsphericalandchromaticaberration,couldhavebeenformedbynaturalselection,
seems,Iconfess,absurdinthehighestdegree."ŠCharlesDarwin,OntheOriginofSpeciesEyeisanextremelycomplexandyetbalancedorganinthehumanbody.Humaneyeisanearly
sphericalorganwhoseprimaryfunctionistoallowforhumanvision.ThevisiblepartoftheeyeappearsasshowninFigure1.1[10].Itiscomparabletoanopticalsystemthatcapturestheimagery
ofasceneandprojectsitontoasensorknownasretinainthebackoftheeye.Texturalpatternofhumaniris(seeFigure1.1)isbelievedtobeuniquetoeachindividual.Thisisexploitedintheofbiometricstorecognizeindividuals.
1.1Biometrics
Passwordsandkeyshavebeenthecornerstoneofauthentication.However,biometricshasmade
inroadsintotheworldofsecureauthenticationandsurveillanceinthe21stcentury[11].ISO/IEC2382-37:2012[12]biometricsasthescienceofautomaticallyrecognizingindividualsbasedontheirbiologicalandbehavioralcharacteristics.Examplesincluderecognizinghumansbasedon
theirface,irisandhandgeometryamongothers.Unlikepasswordsthathavetobe
rememberedorkeys/tokensthathavetobephysicallycarried,biometricsareintrinsicallyassoci-
atedwiththeusersthemselves.AlargestudyofwebpasswordhabitsbyMicrosoft[13]onhalf
millionusersfoundthatanaverageuserhas6.5passwordsandusesthemacrossanaverageof
25accounts.Itgetsincreasinglyhardertocreateandremembernewpasswords.Secureaccessto
physicallocationstypicallyrequireskeysortokenssuchasmagneticcards.Mostlocationsalso1Figure1.1FigureshowingtheexternalanatomyofthehumaneyeintheRGBspectrum.The
focusofthisthesisisontheiris,whichistheannulartexturedstructuresituatedbetweenthepupil
andthesclera.Theirisistypicallyimagedinthenearinfrared(NIR)spectrumandnotinthe
RGBspectrum.
requiretheusertotypeinapasswordbesidesproducingatokenorakey.Theuserscannotbe
authenticatedincasetheuserforgetsthepasswordorforgetstobringthekeys/tokens.Biometrics
eliminatesthesestringentrequirementsandonlyneedstheusertointeractwiththesystem.Agood
biometrictrait[14]isuniversal-allusershaveit,permanent-itisstablethroughthelifetimeof
auser,distinct-itisuniqueacrossmultipleusersandiseasilycollectible.Biometricshasbeen
successfullydeployedinrealworldapplicationsincludingsurveillance,immigrantvat
theportofentry,accesscontrol,ATMsandevenidentifyinglostchildren.Aclassicalbiometricsystemconsistsofabiometricsensor(typicallyacameraimagingthe2biologicaltrait),afeatureextractor,amatcherandadatabasemodule(seeFigure1.2).Abiometric
sensorcapturesthebiometricdatafromtheuser,generallyintheformofadigitalsignal.The
capturedsignalmayhavetobepre-processedtoidentifytheregionofinterestorenhancedto
improveitsquality.Thenafeatureextractortransformsthedataintoanumericalpatternthatcan
laterbeusedforcomparison.Abiometricsysteminpracticeisoperatedinoneofthefollowingthreemodes.EnrollmentInthismode,auserisenrolledbyaddinghis/herfeaturestoadatabaseknownasgallery.Featuresstoredinthedatabasefromtheacquireddigitalsignalarereferredtoasatemplateduringenrollment.Inacooperativeenvironment,anidentityintheformofalabelisassigned
toeachstoredtemplate.Itisalsopossibletohaveasystemwheretheidentityofanenrolled
templateisunknownandlabeledusingnominal[15].Figure1.2Abiometricsystem,duringthe,enrollmentstageaddsatemplatebelongingtoanew
userintotheGallery.
VInthevmode,theuserinteractingwiththesystemclaimsanidentity.Forexample,
considerabiometricsystemdeployedtorecognizeapersonenteringtheUnitedStates.Bob,
whoisalreadyenrolledintotheGallery,isnowinteractingwiththesystemclaimingthathe3isBobandwouldliketoenterthecountry.Thesensorcollectsthebiometricdata(probe)andextractsafeatureset.Invmodeasinglegallerytemplatecorrespondingtothe
claimedidentity,inthiscaseBob,isretrievedfromthegalleryandmatchedagainsttheprobefeatureset.Ifthesimilarityisgreaterthanathresholdvalue,thentheidentityissuccessfully
.Sincethematchingisperformedbetweenoneprobeandonegallerytemplate,itisalsoreferredtoas1:1matching.Thisoperationalmodeistypicallyusedtograntaccessto
securefacilities,verifyingidentityattheportofentry,etc.Asinthecaseofvthefeaturesetisextractedfromthedataacquiredfromauser.
Inthismode,theobtainedfeaturesetismatchedagainstallthetemplatesinthegalleryinordertoretrieveidentitieswhosetemplateshavesimilaritygreaterthanacertainthreshold.
Sincethematchingisperformedbetweenoneprobeandallgallerytemplates,itissometimes
referredtoas1:Nmatching(Nbeingthenumberoftemplatesinthegallery).Forexample,
Tomwhoisapplyingtoenteracertaincountrymayberequiredtopresenthisbiometric
sampleTheextractedfeaturesetisthenmatchedagainstagallery
containingtemplatesofknowncriminalsinordertoapossiblematch.Similarly,the
modecanbeusedinsurveillancescenariostodeterminetheidentityofpeople
ataparticularlocation.AnirisrecognitionsystemhasmultiplecomponentsthataredescribedinSection1.3.Inspiteoftherelativelyhighaccuracyofirisrecognitionsystems,theyarehighlysusceptibletoavariety
ofproblems.Forexample,anacquiredimagethatisoutoffocusmayprobablynotebematched
withitscorrectidentity.ThisincreasestheFalseRejectionRate(FRR)ortheFalseMatchRate
(FMR)[16].Section1.4detailscurrentchallengesintheofirisrecognition.41.2Eyeanatomy
TheocularregioninFigure1.1istheanteriorportionoftheeyeballthatisexternallyvisible.
Thehorizontalcross-sectionoftheanteriorocularregionbroadlyconsistsofthreeregionsnamely
pupil,irisandsclera.Pupilisthedarkholeinthecenteroftheeyeandscleraisthewhitish
portionoftheeye.Irisisthetexturedandcoloredpartoftheeyeenclosedbetweenthepupiland
sclera.Thehumanvisualsystemiscomparabletoanopticalcamerasystemwherethepupilmay
beconsideredasthelens/apertureandtheirisastheaperturestopthatcontrolsthesizeofthe
aperture.Inthecontextofbiometrics,theoculartraitstraditionallyrefertophysicalorbehavioralat-tributesintheeyeglobesuchasiris[17],conjunctivalvasculatureinepisclera[18],retinalvas-
culature[19],OculomotorPlantCharacterstics(OPC)[20]andComplexEyeMovements(CEM)
[21].Periocularregion[22]consistsofupperandlowereyelids,andaedrectangularregion
aroundtheeye.Theuppereyelidisatypeofskinfoldthatisabletostretchoutandcovertheeye
toprotectitfromdust,debrisandsunlight.Periocularregionmayalsocontainother
featuressuchaseyebrowsandmolesontheskininthevicinityoftheeyeregion.
1.2.1Layersofeye
Sincetheiristextureistreatedasabiometrictrait,itisimportanttounderstanditsstructure,
components,physiologyandspectralproperties.Sincetheeyeisa3dimensionalobject,the
imageofaneyeismerelya2dimensionalrepresentationoftheoriginalshape.Figure1.3shows
thesagittalcross-sectionofaneye.Fromanimageacquisitionperspective,thelightfromthesourceencounterscornea,aqueoushumor,irisandlensafterwhichitisprojectedontotheretina.CorneaLightfromanobjectenterstheeyethroughthecorneawhichisatransparentprotective
tissuelayerprotectingtheeyefromexternalworld.Itisthedefensivesystememployed5Figure1.3Sagittarialcross-sectionofiris.Imagepublishedherewithpermissionfrom[1]bytheeye.Itcoverstheentiretyofiriswithanapproximatediameterof11:8mm.Therefractiveindexofcorneallayerisapproximately1.33.Thelightundergoesrefractionsince
itispassingfromairwitharefractiveindexof1.0tocorneallayerwitharefractiveindexof
1.33.Hence,itactsasafocusingelementthatfocusestheincominglightintothepupil.AqueoushumorOncethelightcrossescornea,itentersawaterymediumknownasaqueoushumor.This
regionisalsoreferredtoastheanteriorchamberoftheiris.Aqueoushumorthe
ocularregionandhelpsinmaintainingocularpressurewhiletransportingrequirednutrients
toiristissues.Aqueoushumorconsistsof98%waterandsmallportionsofaminoacids,
electrolytes,ascorbicacid,glutathioneandimmunoglobulins.Spectralpropertiesofaqueous
humourmaybeapproximatedtothatofwatersinceitis98%waterandisusuallytransparent
inthevisibleandnear-infraredspectrum.6IrisAfterthelightpassesthroughaqueoushumor,itencounterstheannularirisregionwitha
holeinthecenter.Theirisactsasadiaphragmbetweentheanteriorandposteriorchamber
oftheeye.Theirisisprimarilydividedintothreelayers-stroma,sphincteranddilator
muscles,andpigmentedepithelium.ThesecomponentsarepictoriallyshowninFigure1.4.Collarette Stroma Dilator muscle Sphinctermuscle Pupillary margin Iris root Posterior pigmented epithelium Figure1.4Differentlayersofiriswhenlookingintosagittalaxis.Theirisgainsitstexturefromitselementsintheanteriorportion,i.e,stromalfaeturessuch
astissues,crypts,anti-crypts,freckles,molesandconcentrationofapigmentation
materialcalledmelanin.Thecoloroftheirisismostlyimpactedbytheconcentrationof
melanininthestroma.Verylowconcentrationsofmelaningivesirisabluishcolor,medium
concentrationgivesitagreen/yellow/hazelcolorandahighconcentrationofmelaningives7irisaverydarkbrowncolor.However,theincidentandimageacquisitionwavelengthalso
playamajorroleintheapparenttextureoftheiris.Irisconsistsofabaselayerofheavilypigmentedcellsknownasposteriorpigmentedep-ithelium.Dilatormusclelinesthetopofthispigmentedepitheliumandisresponsibleforpupildilation.Dilatormusclesareradialandextendfromirisroottopupillaryruff.Their
contractionresultsinpullingthepupillarymargintowardstheirisroottherebydilatingthepupilsize.Sphinctermuscle,ontheotherhand,isacircular(paralleltopupillarymarginor
concentrictopupillaryboundary)musclethatextendsfrompupillarymargintoanimaginary
boundaryknownasCollarette.ItcanbeobservedfromFigure1.5thatthecollaretteistheboundarywherespinchteranddilatormusclesstarttooverlap.Howeveritisimportantto
notethatbothsphincteranddilatormusclesarelocatedbeneaththestromaandhencearenot
visibletothenakedeye.Lens(pupil)Lensisaneartransparentcrystallinebiconvexstructurethatislocatedbehindtheirisand
supportedbysuspensoryligamentwhichisinturnconnectedtotheciliarybody.Partofthe
lensnotcoveredbytheirisisvisualizedasadarkholeknownaspupilintheeyeimage,since
allthelightenteringthelensisabsorbedbythevitreoushumorbehindthelens.The
lensalongwiththecorneaaccountsforallthefocusingpoweroftheeye'sopticalsystem
andhelpstofocustheincominglightontotheretinalwallonthebackoftheeye.Thelight
intensityonretinaisconvertedintoimpulseswhicharethentransmittedtothebrainthrough
theopticalnerve.Theextentoflensexposedtothelightiscontrolledbysphincteranddilator
musclesiniris.1.2.2Apparentiristexture
Asmentionedearlier,theapparentiristextureisdependentonthewavelengthatwhichtheiris
imageisacquired.Letusassumethatthereissufilluminationincidentontheeye.Figure8Figure1.5Locationofsphincteranddilatormusclesthatcontrolpupilconstrictionanddilation,
respectively.
1.6showsthemajorabsorptionelementsonthepathfromtheimageacquisitioncameratotheeye.
Sincethebaseofirisisopaqueandallthelightisabsorbedthoughthelensinthecenter,onlythe
texturepertainingtoanteriorportionofirisiscapturedbythecamera.Table1.1andFigure1.7showthewavelengthsoftheelectromagneticspectrumthatweareinterestedinandtheircorrespondingnames.Visiblespectrumrangesfrom450nmwavelength
denotingbluishcolorto700nmwavelengthdenotingreddishcolors.NearInfra-Red(NIR)covers
wavelengthsfrom700nmto900nmandisusuallyconsideredmonochromatic.ShortWaveInfra-
Red(SWIR)encompasseswavelengthsfrom900nmto1600nm.AbsorptionofliquidwaterandmelaninareshowninFigure1.8(a)and(b).Invisiblespectrum,airandaqueoushumoractaspass-throughwhilethelightisscatteredandfrom
tissuesandmelaninpigmentiniris.Bluecoloredirisescontainveryminuteconcentrationsof9lens Air Cornea Aqeuous humor Vitreous humor 1.00 1.37 1.33 1.42 1.33 Refractive Index Retina opaque opaque Air Water Primary absorption element iris iris Melanin Figure1.6PathfromlightsourcetotheeyeTable1.1Wavelengthrangeforvisible,NIRandSWIRspectrumSpectrumWavelengthrangeVisible400nm-700nmNearInfra-Red700nm-900nmShortWaveInfra-Red900nm-1600nmmelanin,and,hencemostoftheincidentlightisscatteredandinternallyresultingina
bluishappearance(duetoTyndalleffect[23]).Iriseswithhighconcentrationofmelaninappear
darkbrowninvisiblespectrumsincemelaninabsorbsmostoftheincidentillumination.Figure1.9
showsexamplesofthreeirisimageswithvaryinglevelsofmelanincontent.InNIRspectrum,theairandaqueoushumorstillactaspass-throughwhiletheabsorptioncoefofmelanindropsafter700nm.Thisresultsindarkirisesexhibitinggood
texturalpatternsrevealingthemeshworkofcryptsandpossiblepigmentationspots.Figure
1.10showsanimageofadarkbrownirisinvisiblespectrumexhibitingdiscernibletexturalpatterns10Figure1.7Figureshowingwavelengthsofelectro-magneticspectrumrelevanttoirisbiometrics(a)(b)Figure1.8Absorptionspectrumof(a)liquidwater[2]and(b)melanin[3]atdifferent
wavelengthsofelectromagneticspectrum
whenimagedwithaNIRsensor.Figure1.10showsaniristhatisapparentlydevoidoftexturalmorphologywhenimagedinthevisiblespectrumbutthtexhibitsgoodtexturalpatternsintheNIRspectrum.Sinceiristextureis
believedtobeunique,NIRcamerasaretypicallyusedtoacquireirisimagesforbiometricpurposes.11(a)Lowmelanin,Bluish(b)Moderatemelanin,Yellowish(c)Highmelanin,darkbrownFigure1.9Exampleofblue,yellowishanddarkbrownirisimagesthatcontainlow,moderateand
highconcentrationofmelanin,respectively.(b)(c)(a)Darkiris(FalsecolorRGB)(d)(e)Figure1.10Darkirisin(a)imagedat(b)470nm,(c)520nm,and(d)700nmand(e)NIR
wavelengths.
1.2.3Pupildynamics
Iriscontrolsfortheamountofvisiblespectrumlightenteringthepupil(lens).Althoughirismus-clesarecontinuouslyadjustingforthelight,theyareusuallymaintainedatadelicatebalancewith
minimalmovements.Thisstateisknownastherestingstateoftheeye.Howeverexternalfactors
suchasalcoholintake[24],changeinbrightnessandadministeringeyedropdrugs[25]andinter-
nalfactorssuchasdiseaseandstressforceseitherthesphincterordilatormusclestoactivate,and
toconstrictordilatethepupilaccordingly.Figure1.11showsexamplesoffactorsthatpupildilation/constriction.12BeforeAfterMuscleBrightlightonSphincterVisiblelightoffDilatorDrug(Eyedrop)[25]DilatorAlcholConsumption[24]SphincterAlcholConsumption[24]DilatorFigure1.11Examplesoffactorsthesizeofthepupil.1.3Irisbiometricsystem
Inreferencetoanirisbiometricsystem,thebiometricsensoristypicallyaNIRcamerathatacquires
animageofaneyeinthe750nm-850nmwavelength.Itisthenfollowedbyapre-processormodule
thatconsistsofasegmentationprocessthattheirisregion,andanormalizationprocess
thatconvertstheannularregionintoarectangularmatrix.Featureextractormoduleencodestheiris
textureandgeneratesatemplateknownasIrisCodethatconsistsofbinaryvalues.Thesemodules13areshowninFigure1.12.Figure1.12ComponentsofatypicalirisrecognitionsystemBroadly,thecomponentsinFigure1.12maybecategorizedintothefollowingtasks.1.Imageacquisition
IrisimagesaretypicallyacquiredintheNIRspectrum(750nm-850nm).Asdescribedinthe
earliersection,theconcentrationofmelaninpigmentationdeterminestheperceivedcolorof
theirisinthevisiblespectrum.Higherconcentrationsofmelaninresultsindarkercolored
iriseswhileitsabsenceresultsinlighterbluishiriscolors.However,theeffectofmelanin
decreasesintheNIRspectrum[17].Hence,goodtexturalpatternsareobserved,
evenfordarkeririses,intheNIRspectrum.
However,severalworkshavearguedforfeasibilityofirisimageacquisitionintheVisible
spectrum[26][27]andShortWaveInfra-Redspectrum(900nm-1350nm)[28].
Traditionally,irisimageacquisitionrequiredasubjecttopeerintothecameraatcloseprox-
imity.However,recently,therehavebeenseveralsystemsthatareabletoacquiregood14qualityirisimagesfiatadistance"[29][30]upto3metersorfionthemove"[31].Thereisalsoasystemthatisabletocaptureirisimagesasapersondrivesthoughacheckpoint[32].
Thereareotherresearcheffortsthataimtoobtainconsistentlysharpimageswithgoodfocus
byextendingtheviawavefrontcoding[33]andhyper-focalimaging[34].2.Segmentation
Theacquiredimageconsistsoftheocularandperiocularregion.Segmentationistheprocess
ofautomaticallylocalizingtheirisregioninthegiveneyeimage.Aspartofthisprocess,
theinnerpupillaryboundary,theouterlimbicboundaryandthecontoursofupperandlower
eyelidsaredetected.Occludingfactorssuchaseyelashesandspeculararealso
detected.
Therearevariousapproachestothissegmentationtask.Daugman,in[17],proposedan
integro-differentialoperatorthataimstoaboundarythathasamaximumcumulative
radialimagegradient.TheIntegro-differentialoperatorisgivenbymax(r;xo;yo)jGs(r)¶¶rIr;xo;yoI(x;y)2prdsj:Thealgorithmcomputesthecumulativeradialimagegradientateverypixelinthecircumfer-
enceofacirclewithaedsizeradius.Thisprocessisrepeatedformultipleradiusvalues,r.Thecirclethatresultsinthemaximumcumulativevalueisdetermined.Thiscancorrespond
totheinnerorouterboundaryoftheiris.
Wildesin[35]detectstheedgesintheimagesandconvertstheinputimageintoabinary
edgeimage.ThenacircularHoughtransformisusedtoidentifycircularboundaries.Fora
edacquisitiondistance,upperandlowerlimitscanbesetfortheouterboundaryradius.
Theselimitsareusedtoeliminatefalsepositivesandselectthecorrectirisboundary.The
regioninsidetheouterboundaryisthensearchedtotheinnerpupillaryboundary.Line
Houghtransformisusedtodetectupperandlowereyelids[36][4].15However,recentworkinirissegmentationhasfocusedonremovingtheassumptionofcir-
cularboundaries,sincethelimbicandpupillaryboundaryarenottypicallycircularunder
non-idealconditions.ZuoandSchmidin[37]approximatedtheirisandpupilboundaries
withmorerelaxedellipses.ShahandRoss[38]furtherremovedtheconstraintsbyde-
tectingthepupilbythresholdingandthenusingasnakelikegeodesicactivecontourto
thelimbicboundary.Thereareothersimilarworkthatrelyontheprincipleofactivecontours
[39][40],althoughthedetectionofpupilisstillperformedusingbasicthresholdingfollowed
bybinarymorphologicaloperations,sinceitslocationisneededtoinitiatetheactivecontour.
Othermethodsinvolveclassifyingthepixelsbasedontheirtexturalcontent.Broussardet
al[41]usedaneuralnettoclassifyeachpixelasirisornon-iris.Thesemethodsinvolve
extensivetrainingtobuildmodelsthatlearnthedifferencebetweentrueirispixelsandnon-
irispixels.Heetal[42]usedatrainedAdaBoostdetectortorapidlylocalizetheirisregion
(rectangularboundingbox).3.Normalization
Normalizationistheprocessofunwrappingtheannularirisregionintoaedsizerectangu-
largrid.Normalizationisexpectedtoaccountfortheiristexturedeformationduetovarying
pupilsize.Normalizationisassumedtoresultinverysimilarrectangularimagesevenifthe
imagesofthesameeyearecapturedwithdifferentpupilsizes.However,recentworkshas
showntheinadequacyofthisassumption.Mostmethodsareeitherbasedonorarevariants
ofDaugman'srubbersheetmodel[17].Thisstepisoptionalsincetherearemethodsthatperformimagematchingontheoriginalimagesthemselvessuchas[43],thatusedsimilarity
ofdescriptorsatlocalinterestpoints,and[44],thatusedclassicSIFTdescriptortomatch
irisimages.
Therubbersheetmodelmapseachpixel(x;y)intheirisregiontoapoint(r;q)inthe16rectangularregionusingthefollowingmappingfunction.I(x(r;q);y(r;q))!I(r;q)where,x(r;q)=(1r)xp(q)+rxl(q)y(r;q)=(1r)yp(q)+ryl(q):Here,xp(q);yp(q)andxl(q);yl(q)areasetofpupillaryandlimbicboundarypoints.Theformulacanbeinterpretedasfollows.TheannularregionissampledatRregularintervalsalongtheradialdirectionataedangularvalue.Thesampledpointsareassembledalonga
singlecolumnofthenormalizedimage.Thisisrepeatedacrossmultipleangulardirectionsto
populateothercolumnsinthenormalizedimage1.14.Similarly,anormalizedmaskisalso
generatedtodenotethenon-irispixelsthatcorrespondtotheeyelids,specular
eyelids,etc.4.Patternrepresentationandmatching
Sincetheiristextureisbelievedtobeunique,thereareseveraltexturerepresentationmethods
andcorrespondingdistancemeasurestomatchtwoirisimages.Classicalmethodinvolves
convolvingthenormalizedimagewithabankofcomplexGaboroftheformG(r;q)=eiw(qqo)er(rro)2a2e(qqo)2b2;where,roandqodenotetheradialandangularbandwidthofthe2-DGaborFigure1.13showsrealandimaginarypartsofaGabor.
Therealpartoftheresultingoutputisadjustedtohavezeromean.Thentheadjustedreal
partandthecomplexpartarebinarizeddependingonthesignoftheresponse.Positive
valueisdenotedas1andnegativeoutputisdenotedasazero.Hence,foreachpixelinthe
normalizedimage,twobitsaregeneratedusingone.Thebinaryrepresentationof17(a)(b)Figure1.13(a)Realpartand(b)ImaginarypartofaGaborthenormalizedirisimageisreferredtoasIrisCode.IrisCodesCAandCBwithcorrespondingmasksMAandMBarecomparedusingafractionalHammingdistance:HD=(CANCB)T(MATMB)kMATMBk:Inprinciple,thisvaluemayrangefrom0(completematch)to1(completemis-match).In
practice,theimpostorscoreshaveameanof0.5sincetheprobabilityoftwocompletely
randombit-streamsmatchingisaround0.5.
OthermethodsthatusesimilarapproachesincludeBolesandBoashash[45]thatusezero
crossingof1Dwavelettransform,Chouetal.[46]thatusesLaplacianofGaussian
Rocheetal.[47]thatusezerocrossingsofdyadicwavelettransform.
Therearealsomethodsthatuseseigen-irisapproachthatattempttoextractbasisfunctionsandrepresenttheinputimageasacombinationofthesebasisfunctions.Examplesinclude
methodsbyDorairajetal.[48]whousedPCAandICAontheentireregion,Huanget
al.[49]whoappliedICAonsmallwindows,Maetal.[50]whousedGaborin
conjunctionwithFisher'sLDAtodiscriminatebetweenirisimages.
OthertexturaldescriptorsincludeGLCM(GrayLevelco-occurrenceMatrix)thatwasused
byChenetal.[51]inwhichtheycomputeda3-Dco-occurrencematrixinsteadoftheclassic
pairwiseco-occurrencematrix.LBP(LocalBinaryPatterns)isalsousedtodenotetextural18patternsinnon-overlappingblocksinthenormalizedimage,andablocklevelsimilarity
measureisusedtocomputedistancemeasure.
Figure1.14showstheoutputsofsegmentation,normalizationandencodingmodulesona
sampleirisimage.(a)(b)(c)(d)(e)Figure1.14(a)OriginalacquiredImage(b)Segmentationoutput(c)Normalizedimage(d)
Correspondingmaskimage(e)IrisCodegeneratedbyencodingthenormalizedimageusing
Masek'smethod.[4]
Otheroptionalmodulesincludequalitycheckertoaccept/rejecttheacquiredimagesbasedonthe
qualityoftheacquiredimage,andapre-processingmodulethatenhancesthequalityofeitherthe
acquiredimagesorthesegmentediristexture.191.4ChallengesinIrisrecognition
Therearemultiplefactorsthattheperformanceofanirisrecognitionsystem.Mostof
themareduetointeractionbetweensensorandtheuser,whileothersareduetothecharacteristics
oftheeyeandthechoiceofimageprocessingmethods.Itmaybenotedthatirisrecognition
systemshaveaverylowFalseMatchRate(FMR)providedsufnumberofbitsarematched
(lowocclusion).Hence,thesechallengingfactorsincreasetheFalseNon-MatchRate(FNMR)i.e,
theyresultinfailureofsuccessfullymatchingimagesofthesameeyeacquiredatdifferenttimes.
Alistofsuchfactorsispresentedbelow.1.UserInteractionandAmbientFactors(a)IlluminationPoorilluminationisnotamajorconcernunlesstheilluminationintensityisconsider-
ablylowthatresultsinthesensorregisteringdarknoiseinsteadofactualtexture.How-
ever,non-uniformilluminationisaveryseriouschallenge.Ontheotherend,strongil-
luminatorscanresultinlargespecularwhichmightimpactiristextureand,
insomecases,affectsegmentationaccuracy.Figure1.15showsexamplesofpoorly
illuminatedimages.(b)OcclusioniEyelidsSometimestheusersmaynothavetheireyescompletelyopen,seeFigure1.15,
thatwouldresultinimageswhere,irisisoccludedbytheeyelids.Itreducesthe
numberofirispixelstherebyreducingthediscriminativepoweroftheacquired
image.iiEyelashesSomeindividualsmayprefertohavelonganddarkeyelashes[52].Sucheyelashes20canoccludeapartoftheiris.Oneofthemajorchallengehereistodetectthe
eyelashesinordertoexcludethemduringthematchingstage.iiiGlassesAlthoughclearglassesarenotbelievedtoimpactiristexture,itbringsinadditional
challengessuchasspecularandframeocclusions.ivContactlensCertaintypesofcontactlensessuchashardlenses,markedlensesandtheatrical
patterncontactlensesareshowntodegradeirisperformancebyaconsiderable
margin.However,itispossibletodetectthepresenceofsuchcontactlenses.(c)FocusIrisrecognitionsystemsexpectawellfocusedimagethathashighfrequencycontent
init.Stronglydefocusedimagessmoothoutthetextureandtheresultingencoded
informationwouldcorrespondtothestateofthesensoratthetimeofcapturerather
thantheoriginaltexture[17].However,itiseasytorejectsuchkindofimagesby
computingthefocusmeasurerapidlyinrealtimeandretainingonlyin-focusimages.(d)MotionblurIrisislocatedonacontinuouslymovingorganknownastheeyeballwhichisinturn
placedinanothermovingobject-thehead.Hence,itispossiblethattheimagespro-
curedbythecameraexhibitaamountofmotionblur.(e)ImageresolutionTypicalirisimageacquisitionsystemsrequiretheusertointeractwiththecameraat
closeproximity.Itensuresgoodimagequalityintermsoffocus,bluranduniform
illumination.Butanothermajorchallengeassociatedwithlargestandoffdistanceis
poorimageresolution.Itisrecommendedtohaveatleast200pixelsacrosstheiris
diameter[53]toachievegoodirisrecognitionperformance.(f)Off-axisirisimage21(a)(b)(c)(d)(e)(f)Figure1.15Examplesofnonidealirisimages.(a)and(b)Non-uniformillumination,(c)and(d)
Eyelidocclusion,(e)Eyelashocclusion,(f)Motionblur.Irisrecognitionsystemsrequirethecapturedirisimagetobefrontal,i.e,theeyehasto
bestaringdirectlyintothecamerainlinewithitsopticalaxis.Otherwise,theacquired
imagewouldbedeviatedfromtheopticalaxisintheroll,yawandpitchdirections.
Figure1.16showsexamplesoffewoff-axisimages.Off-axisimagesmwhencompared(a)(b)(c)Figure1.16Examplesofoff-axisirisimages.againstfrontalenrolledimages,wouldnotyieldthesamenormalizedimagenorcan
becompareddirectlysincethereisanaftransformationinvolved.Althoughsucha
transformationmatrixmaybeestimated[16],itmaynotbecompleteandreconstruction22ofonesetofimagesfromtheotherisnot2.SensorItispossibleforaniristobeenrolledusingonecamerasensormodelbutrecognizedusing
imagesacquiredbyadifferentcamerasensormodel.Bowyeretal.[54]observedthat
althoughthenon-matchdistributionisstable,thematchscoredistributionsareadversely
impacted.3.ImagecompressionBiometricdatamaybestoreddigitallyonpassports.Itisalsosometimesnecessarytostore
theoriginalimageratherthantheIrisCodetemplate.Insomeapplications,thisimagehasto
bestoredinlimitedspace.Forexample,theRegisteredTravellerInter-operabilityConsor-
tium(RTIC)[55]allocatesonly4000bytespereye.Atypicalgrayscaleirisimageofsize
640480has307;200bytesofdatathathastobecompressedto4000bytesbyascaleof76:8.RakshitandMonro[56]showedthatthenormalizedorfiunwapped"irisimagecouldbecompressedto2560bytesandDaugmanandDowning[57]showedthattheoriginaliris
image(innativeimagedomain)couldbecompressedtoaslowas2000byteswithoutsub-
stantiallyimpactingtherecognitionperformance.4.EyediseasesEyediseasescanadverselyimpactirisrecognition[58][59][60]sincetheymaydeformthe
observediristexture,distortpupilshapeorimpacteyecolor.Figure1.17showsexamplesof
irisimagesexhibitingeyediseases.Itcanbeobservedthatinsomeoftheimages,contours
oftheirisboundariesaredrasticallyalteredandtexturalabnormalitiesareinduced.5.IrisstabilityHumanirisstartsformingfromthethirdmonthofgestation.Theconstituentpartsofthe
iriscontinuetogrowandstabilizeafter8monthsofconception.However,thepigmentation
continuestogrowafterbirthuntilthesecondyear.However,therearemanytheoriesfor23(a)(b)(c)Figure1.17Fewexamplesofeyediseasesthatimpactirisrecognition(a)Polycoria-Multiple
pupilopenings(b)Coloboma-Teariniris(c)Severecataract-Thickeningoflens(looses
transparency).AlthoughtheimagesareshowninRGB,someofthesediseasescanalsoimpact
theNIRimages.predictingtheeyecolorgivenfamilyhistoryofeyecolors[61].Itiscommonlybelieved
thattheiristextureremainsrelativelystable(exceptinthecaseoftheeyediseases)aftertwo
yearssincebirth.HoweverFenkerandBowyer[62]havepresentedevidenceofmatchscore
degradationwhencomparingimagesofthesameeyetakentwoyearsapartusingthesame
camera.Thisphenomenonisreferredtoasirisaging.Itmustbenotedthatirisagingmaybe,inpart,duetothelimitationsofirisrecognitionalgorithmandintra-classvariationdueto
variationsinpupilsizeandimagingconditionssuchasblur,focusandgazedirectionsacross
imagingsessions.6.PupildilationPupilrespondstothestrengthoflight(invisiblespectrum)enteringtheeye.Itconstricts
inbrighterlighttoprotecttheretinaanddilatesindarkerenvironmentstoallowformore
lighttoentertheeye.Daugman'srubbersheetmodelfornormalizingtheirisimage[17]is
believedtoaccountforchangesinpupilsizeacrossdifferentlightinglevelsandimagesize.
Howeverrecentresearch[6]hasshownthatextremevariationinpupilsizewouldincrease
theHammingdistancebetweensamplesofthesameeyeresultinginfalsenon-matches.7.Multi-spectralmatchingAlthoughtheirisisimagedintheNIRspectrum,therearepracticaltobeableto24performirisrecognitioninvisiblespectrumespeciallyduetotheadventofsmartphonesthat
typicallycaptureimagesinthevisiblespectrum.Also,Rossetalin[28]haveshownthe
feasibilityofperformingirisrecognitioninwavelengthsrangingfrom900nmto1350nm.
ThesewavelengthsareconsideredtobepartoftheShortWaveInfraRed(SWIR)spectrum.
HumaneyeisnotabletosensethesewavelengthsandastrongilluminatorinSWIRband
wouldbeinvisibletoahumanobserver,makingitviableforuseincovertaswellasnighttime
environments.Itisalsosometimesrequiredtomatchanirisimageacquiredineithervisible
orSWIRbandagainstanNIRtemplatestoredinthedatabase.
Themajorlimitingfactorstoperformintraspectralorcross-spectralmatchingareŁLackoftexturalcontentindarkeririswhenimagedinvisiblespectrum.
ŁSpecularectionsinvisiblespectrumduetothetearonthecorneallayer.
ŁDifferentialresponseofirisconstituentsatdifferentwavelengths.1.5Objectivesofthiswork
Thisworkfocusesononeofthemajorchallengesfacingirisrecognition,namelypupildilation.Theadverseimpactofpupildilationisstudiedandasimpleyeteffectivesolutionisproposedto
improvetheperformanceofirisrecognitionwhentheinputimagesexhibitlargedifferenceiniris
sizes.25CHAPTER2MOTIVATIONANDPREVIOUSWORKIrisisacomplexstructureinthehumaneyethathasveryinterestingelasticproperties.When
thelightincidentontheeyeisvaried,musclesintheiriscontractorexpandtoallowforlessor
morelightintotheeyetobetterperceivethescenewhilstprotectingtheretinaatthesametime.
Interestingly,theirismusclesrevertexactlytotheiroldpositionafteraperturbation[63].
2.1Motivation
Duringthenormalizationstage,mostirisrecognitionalgorithmsunwraptheirisintoapseudo-
polarcoordinaterectangleusingDaugman'srubbersheetmodel[17]bysamplingtheirisregion
uniformlyalongtheradialandangulardirections.Thistransformationisbelievedtoaccountfor
changesinirissizeduetoitscompressionordilation.However,uponsimplevisualobservation,
itisevidentthatirisundergoesacomplexnon-lineardeformationduringpupilconstrictionor
dilation.Itiswelldocumentedthatextremepupildilationaffectsthematchscorebetweentwo
irisimages[6].Largerthepupilsizedifferencebetweentwoimagesofthesameiris,largeristhe
Hammingdistance.Figure2.1shows(a)aneyeimageand(b)it'scorrespondingnormalizediris
image[5].Whenthepupildilatesfrom(a)to(c),theirisregioniscompressedinanon-linear
fashionasshownin(d).Close-upofregionsinFigure2.1(b)and(d)showsthatthesehighlighted
regionsdonotalignwellwitheachother.Hollingsworthetal.[6]showedthatalargedifferenceinpupilsizebetweentwoimagesresultsinalargegenuinedissimilarityscore.Figure2.2showstwoirisimageswithdifferentpupil-to-iris
radiusratiovalues.Ifthedilationratioisastheratioofpupilradiustoirisradius,thena
smallervalueofpupildilationratioindicatesalargeririsregionwithasmallerpupilsizerelativeto
theirisradiusandalargerdilationratioindicatesalargerpupilsizewithrelativelylessirisregion.
Itisnotuncommontoirisimagesthathavedilationratiosaslowas0.2andashighas0.8[6].26(a)(b)(c)(d)Figure2.1(a)and(b)Irisimagewithmoderatepupilsizeandthecorrespondingnormalizediris
image.(c)and(d)Irisimagewithlargepupilsizeandthecorrespondingnormalizedirisimage.
Highlightedregionsin(c)and(d)donotaligncorrectly.Imagesfrom[5].Figure2.2Irisimageswithdilationratiosof(a)0:3478and(b)0:6545.Imagesfrom[6].Ineffect,theeyeimageacquiredatdifferenttimescanexhibitalargevariationindilationratio,therebyincreasingthepossibilityoffalsenon-matches,wheretheuserisfailedtobeHence,thereisaneedtoaccountforthevariationsiniristexturetobettermatchtwoirisimageswithlargepupilsizevariation.272.2Previouswork
Thepreviousworkonthistopicmaybebroadlydividedintothreecategoriesbasedontheirend
goals.Thelineofworktriedtomodelthedynamicsofirisdeformationbyderivingatheo-
reticalmodeltounderstandthedeformationprocess.Thesecondlineofworkonlyemphasizedon
improvingtheirismatchingperformanceinpresenceofpupildilationwithoutnecessarilymodel-
ingthebiologicalbasis.Thethirdcategoryofworkonlyfocusedondocumentingtheeffectsof
pupildilation.Thefollowingarethreedeformationmodelsproposedintheliteratureinchronologicalorder.1.Minimumwearandtearmodel;
2.Empiricalmodel;
3.Mechanicalstrainmodel.2.2.1Minimumwearandtearmodel
Rohen[7]wasthetoproposeastructureforcollagenousiniris.Figure2.3showsthe
structureproposedby[7]thatconsistsoforthogonalsetof(clockwiseandanti-clockwise)
thatconnectthepupilboundarytotheouteririsboundary.Rohenalsoobservedthatthese
areinterwovenwithbloodvesselsandothercomponentsoftheiris.
2.2.1.1Wyatt2000
Wyatt[8]providedamathematicalframeworkforthismeshworkthatminimizeswearandtearofirismusclesduetoconstrictionordilation.Thereareadditionalconstraintsthathavetobe
forbetterapplicationofthismodelforirisdeformation.Forexample,pointsontheirisshouldnot
rotatetoomucharoundthecenterofpupilasthepupildiameterincreases.Secondly,the
arcsinthemeshworkmustnothaverelativeslipatanygivenlocation.Theconditionslaidbythe28Figure2.3IrismeshworkproposedbyRohen[7].Imagefrom[7].constraintsaremetwhenpointsintheirisregionareassumedtomoveonlyinradialdirectionas
pupildiametervaries.WyattmodeledlineardeformationofirisaccordingtothefollowingformulaR(q;qo;p)=R(q;qo;pref)ropropref+ropprefropref:Ristheradiusasafunctionofpolarcoordinateq,thepolarangletraversedbyasinglefrompupillarymargintotheirisrootqo,andpupildiameterp;themeshworkisinitializedwiththepupildiameterequaltopref.Figure2.4showsapictorialrepresentationforqo.ThemeshworkwasrepresentedusingasimplelogarithmicspiraloftheformR=pRopqqo:Aftersolvingforlogarithmicspirals,additionaldeviationwasallowedintheformofa20-term
polynomialinqtoaccountfornonlineardeformation.Anoptimumcurvewasfoundforq=100asshowninFigure2.5.Thenonlinearstretchofirisismodeledasthesumofalinearstretchand29Figure2.4qoistheanglebetweenstartingpointofthearconpupillaryboundaryandendingpointonlimbicboundary.Figure2.5OptimumarcsderivedbyWyatt[8]forq=100,andpupildiameter1.5,4.0and7.0mm
anonlineardeviation.R=Rlinear+DR(p;r):where,RlinearisthesolutionofthelineardeformationmodelandDRistheadditionaldisplace-mentofapointontheirisregionafterthelinearstretch.DRisapproximatedusinga6thorderpolynomial.302.2.1.2YuanandShi2005
YuanandShi[9]leveragedtheideaofmeshworkasanddescribedamodelforestimating
thelocationofapointintheirisregionafterdeformation.Semi-circulararcsareconstructedas
showninFigure2.6.FromtheP0isthereferencepupilboundarywhichisdeformedtotheFigure2.6NormalizationmodelproposedbyYuanandShi.Imagefrom[9].currentboundarymarkedasP.I0istheirisrootboundarywhichisassumedtoremained.Inthisimplementation,theanglebetweenanyPandit'scorrespondingI0isp=2.Thearcsbeforeandafterdeformationaremodeledassectorsofcircles.GivenalocationA0intheirisregionofthereferenceimage,it'scorrespondinglocationAafterdeformationcanbeeasilyderivedasafunctionofthepoint'slocationwithrespecttothepupilcenter.Theassumptionsmadeinthismodelare:Łthepupillaryandlimbicboundariesareapproximatedasconcentriccircles;
Łmarginofthepupil(boundary)doesnotrotateand
Łshapeofpupilremainsroughlycircularduringdilationorconstriction.31Fromthemodel,itisevidentthatpointsclosertothepupilboundaryaredisplacedbyalarge
distance,whilepointsclosertotheirisroot(limbicboundary)arenotdisplacedasmuch.This
introducesanonlinearityindisplacementmagnitudesforpointsintheirisregionasafunctionof
theirdistancefromthepupillaryboundary.Aparameterlisasl=rR;where,ristheradiusofpupilandRistheradiusoftheouteririsboundary.Asinthepreviousmodel,aedpupilradiusischosenasthereferenceusingtheformularref=lrefR.Thedeformationmodelisusedtodeformthegivenirisasitspupilradiuschangesfromrtorref.Oncethegivenirisimageisdeformedtomatchpupilradiusequaltorref,thenitislinearlymappedtoapseudo-polarrectangulargridusingDaugman'smethod[17]forfurtherencodingandmatching.
2.2.1.3Weietal.2007
ThemodelproposedbyWeietal.[5]followsalongthesamelinesasWaytt[8]bymodelingthe
nonlinearstretchofpointsinirisregionsassumofalinearstretchandadeviation.Thisdeviation
ismodeledasafunctionofthecurrentpupilradius,pandposition,r:Rnonlinear=Rlinear+DR(p;r):WhileWaytt[8]approximatedthedeviationvalueasa6thorderpolynomialinq,Weietal.computedthedeviationvaluesusingstatisticalmeasuresofatrainingset.Astheirisradiusmaydifferslightlydependingontherelativepositionoftheeyetothecameraduringimageacquisition,aconsistentparametercalledirisdeformationfactorTisasT=RpRi;where,RpandRiareradiusofthepupilboundaryandtheirisrootboundary,respectively.DRisthenmodeledasafunctionofRlinearandT,Rnonlinear=Rlinear+DR(Rlinear;T):32Thisirisdeformationfactor,T,issameasthedilationratioin[6].AreferencebandforT,namely[Ts;Tl],ischosenandthedeformationmodelisappliedwhenthevalueofTisoutsidethisband.ThepupilisdilatedforT>Tl,andthepupilisconstrictedforT<Ts.ThedeviationfromthelinearstretchisafactorofhowfarTisfromthereferenceband.ThedeviationfromlinearstretchisformulatedasDR=CF(Rlinear);where,C=(Ts+Tl2T)andF(Rlinear)isafunctionofthelinearstretch,Rlinear.Here,Cdeter-minesthestrengthofnonlineardeformation.F(Rlinear)islearntusingatrainingsetof600irisimagesfrom60subjectswith10sampleseach,thatwereobtainedatgraduallyvaryingillumination.Asetofpointsintheirisregionare
manuallymarkedandtrackedacrossthe10images.Thesetofpointsaredividedintothreeregions,
fPing,fPmidgandfPoutg,basedontheirproximitytothepupilboundaryusingnearestneighborclustering.NonlinearstretchiscomputedforthesethreeregionsanddeviationDRisderivedforallthethreeregions.TheplotteddeviationsfortheseregionsareapproximatedusingGaussian
distributions.DR=CF(Rlinear)=CN(m;s2)=C1p2psexp((Rlinearm)22s2):ParameterCistheirisdeformationfactorthatcanbeestimatedfromtheirisimage.(m;s)canbeestimatedfromplotsofdeviations.
2.2.23-Danatomicalmodel
2.2.2.1Francoisetal.2007
Thisisnotanirisdeformationmodelbutananatomicalrepresentationoftheirisstructurethat
couldpotentiallybeusedtomodelirisdeformation.Irisisa3-Dentityconsistingofstructures33atdifferentdepthsfromthecornealplane.Theincidentlightisrefractedontothesestructures
anda2-Dprojectionontherecordingcameraisvisualizedasanimage.Francoisetal.[64]
proposedamethodtorecoverstructureofirisfromasinglephotograph.Arepresentationknown
asSubsurfaceTextureMapping[65]isusedtodescribethemorphologicalreliefofthehumaniris.
Thenarefractivefunctionisalsopresentedtoaccountforrefractionatcornealsurface.
2.2.2.2Clarketal.2012
ThemechanicalmodelproposedbyClarketal.[66]considersirisasamaterialthatisactedupon
byinternalmechanicalforces,andsubsequentdeformationismodeledintermsofmechanical
strain,stressandmaterialpropertiesofiris.Amathematicalmodelisderivedusingbiomechanics
ofiristocharacterizethenonlineardeformation.Theirisisapproximatedasathincylindricalshellwithnegligiblethicknessinthezdirectionandthestructurereducestoathinplateandcanbemodeledintermsofpolarcoordinatesrandq.Thedisplacementofapointintheirisregion,whenthepupilradiuschangesfromsomeinitialvaluetosomevalue,isrepresentedasu(r;q).~u(r;q)=ur~r+uq~q:Cauchy-Eulerequations[67]forthinplatesareusedforstrainequilibriumconditionswhilea
separatesetofstressequilibriumconditionsarealsoderived.Additionalassumptionofnegligibleangulardisplacement,uqand¶¶q,ismadebasedontheobservationthatpupillaryresponsecausesaxisymmetricloadontheirismusclesandthatirismus-
clesareequallydistributedacrosstheirisregion.Theseassumptionsalsoleadtothe
ofshearstress[68].Now,thedisplacementvectorbecomesu=ur~r.Thereducedequilibriumconditionsbasedonadditionalassumptionsareasfollows:34ŁForstrain:er=dudr12dudr2:(2.1)eq=ur12ur2:(2.2)ŁForstress:dsrdr+srsqr=0;(2.3)where,erandeqarenormalstrains;srandsqarenormalstresses,respectively.Relationbetweenthestrainvectorandthestressvectoriscomputedassumingtheirismaterialtobeorthotropicthatcanbedeformedonlyintwoorthogonaldirections(randqinthiscase).er=srErnrqEqsq;(2.4)eq=nqrErsr+sqEq;(2.5)where,ErandEqareYoung'smoduliofelasticityforiris;nrqandnqrarePoisson'sratiooftheirismaterialintheazimuthalandradialdirections,respectively.Thesymmetrypropertyfor
orthotropicmaterialstatesthatnqrEr=nrqEq:(2.6)Substitutingequations2.4and2.6inequations2.1and2.3givesrisetoamasterdifferentialequa-
tionoftheform,u00+u0rzur2(1nz)2ru02(n1)z2r(ur)212ddr(u0)2nz2ddrur2=0;(2.7)where,()0isregulardifferentiationw.r.tr;z=EqErandn=nqr.Equation2.7issolvedasboundaryvalueproblemwithconditionsu(pupil_boundary)=candu(limbic_boundary)=0:35Here,cisthedifferenceinpupilradiusbetweentheinitialandAnotherkeyas-sumptioninboundaryvaluesisthatirisremainsedatthelimbicboundary.Thatis,displacement
ofpointsonthelimbicboundaryiszero.Themasterdifferentialequationissolvedusingthe
elementmethodandnumericalresultsshowanonlineardeformationoftheirisregionwithvaria-
tioninpupilsize.Theyalsoshowthesolutionforthemasterequationalongwithnumericalresults
andsimulationwhilstassumingthattheirisregionisisotropic.Isotropicdeformationisconsidered
tobeaspecialcaseoforthotropicdeformationwhenEr=Eq.Intheirworkitisobservedthatlineardeformationisagoodapproximationfornonlineardeformationincaseofsmallerchanges
inpupilsize,butastrongnonlineardeformationisclearlyevidentwhenthepupilsizechangesby
alargemagnitude.
2.2.3Gejjietal.2015
Genjjiet.alin[69]andClarket.alin[70]studiedtheresponseofpupiltolight,alsoknownaspupil
lightx(PLR),inthenearinfra-red(NIR)spectrumusingabiologicalmodel.
2.2.4OtherDeformationmodels
Alltheaforementionedmodelsassumethattheirisisa2-Dstructureandthatangulardisplacement
isnegligible.However,irisisinfacta3-Dstructureandtheimageacquiredisonlya2-Dprojection
ofthetexture.Therefore,thereisaneedtomodeltheirisdeformationasa3-dimensionalobject.
Additionalstructuressuchascontractionfurrowsbecomeevidentwhentheirisiscompressed.It
isstillverycomplextomodeltheirisasa3-Ddeformableobjectsincenoinformationisavailable
inthe3rddimension.Thereareseveraldeformationmodelsintheliteraturethatcanpotentiallybeusedtomodeltheirisdeformationsuchas(a)dynamicmodelingoflocalandglobaldeformation
[71],(b)utilizingprinciplesofdeformationsofelasticmaterialfromcontinuummechanics[72],
(c)modelingusingsplinesandtheirvariants[73],and(d)developingmodelsusingedanchor
pointsaroundwhichdeformationoccurs.36Severalpapershavebeenpublishedthatdescribetheimprovementofperformanceofirismatchingbetweeneyeimagesofvaryingpupilsizes.Forexample,YuanandShi[9]usedthe
minimumwearandtearmodeltoderiveanequationthatpredictstheexactlocationofapointin
theirisregionafterdilation.Inanotherwork,Weietal.[5]approximatedthenon-linearterm
inWyatt'smodelDR(p;r)usingaGaussiandistributionthatisinturnlearnedfromatrainingset.Thorntonetal.[74]dividedthenormalizedirisintoasetofnon-overlappingblocks,andcomputed
transformationparametersbetweencorrespondingblocksinthetargetimage.Thentheposterior
probabilityoftheseparametersismaximizediterativelyresultingintheoptimaldeformationpa-
rameterset.Thisinformationisusedtocomputeblock-wisesimilaritymetricsthatareaveraged
toproduceascore.Tomeo-Reyeset.al[75]usedthebio-mechanicaliristissuemodelusedin
[66]topredictthedisplacementofapointintheirisatagivendilationlevelandusedthisinthe
normalizationprocess.TheytestedtheirtechniqueontheWVU-PLRdatasettoshowa
improvementinmatchingperformanceespeciallywhencomparingirisimageswithlargevariation
inpupilsizes.Pamplona[25]collectedextremelydilatedimagesofafeweyesbyadministeringmydriaticdrugsthatdilatedthepupil.pointsintheirisregionwerethenmanuallyannotatedand
trackedacrosstheimages.Itwasobservedthatpointsaredisplacedpredominantlyintheradial
directionandstructuressuchascryptsdeformintheangulardirection.Thereareseveralpapers
[6][76][77]thatdemonstratetheadverseimpactofpupildilationonirismatchingperformance.
2.3Bitmatching
TherehavebeenothermethodsdevelopedthatexploitthecharacteristicsoftheIrisCodetoimprove
theperformanceofanirisrecognitionmatcher.Hollingsworthetal.[78]usedamatchingscheme
whereonlybestbitsinanIrisCodeareused.Bestbitsarechosenbasedontheirconsistencyacross
differentsamplesofthesameeye.Rathgebetal.[79]employedaselectivebitmatchingschemeby
comparingonlythemostconsistentbitsinanIrisCode.Theseconsistentbitsareobtainedbyusing37differentfeatureextractors.Inotherworks,Rathgebetal.[80]proposedanewdistancemeasure
basedonHammingdistancevaluesthataregeneratedbyshiftingoneIrisCodewithrespecttothe
otheratmultipleoffsets.InSLIC[81],IrisCodesarematchedonerowatatime,therebydecreasing
thediscriminatorypotentialofIrisCodesthataretypicallymatchedintheirentiretybutresulting
inbettermatchspeeds.38CHAPTER3COLLECTIONOFDATABASE3.1Motivation
Majordrawbacksofthepreviousworkintheliteraturetoaddresstheproblemofpupildilation
areasfollows.1)Theoreticalmodelsarenotempiricallyvalidated;2)Softwaresolutionsrequire
alterationstoexistingsystems;and3)Datasetsusedinpreviousresearchdonotsystem-
aticallymeasuretheimpactofpupildilationonirismatching.Previouslydemonstratedeffectsof
pupildynamicsweretestedongenericdatasetsthatwerenotacquiredforstudyingthe
effectofpupildilation.Asnotedearlier,thereareseveralothercontributingfactorssuchasfocus,
illuminationchangesandblurthatcanimpactrecognitionaccuracy.Inourwork,thesefactors
areovercomebyacquiringadatasetinhighlycontrolledilluminationconditionsanddistancesasdescribedinthefollowingsection.
3.2Dataacquisitionprotocol
VideosarecapturedwithaRedlake(DuncanTech)MS3100multispectralcameraatroughly8
frames/sandsavedasasequenceofstillimages.Thecameraisattachedtothemobilearmofan
ophthalmologist'sslitlampandconnectedtoanEpixframegrabber.Anannularringlighted
by2NIRLEDs(810nm)isplacedinfrontofthecameraandisconnectedviaanopticguide
toaStelarNetlightsource(avoltageregulatorandatungsten-kryptonbulbwithabroadspectrum
of300nmto1700nm).ThetwoLEDsareusedforanevenilluminationoftheeyewhilecamera
isfocusedpriortodatacollection.Withthechinonthechinrestandgazingintothecamera,the
participantisgiventimetoadjusttothedarkness.Withcamerainfocus,therecordingisstarted.
After10seconds,theon/offbuttononthelightsourcepanelisturnedon,thelightisdirectedto
theeyethroughtheannularringforanadditional10secondsintervaloftime,afterwhichthelight39isturnedoff.Thevideorecordingisstoppedfollowing10secondsofdarkness.TheNIRLEDs
areonforthedurationoftherecording.Thevideocapturesthepupildynamics:theconstriction
ofthepupilwhentheeyeisexposedtotheoflightandthedilationofthepupilwhenthe
eyeadaptstothedarkness.Figure3.1depictsthevariationofthevoltageonthetungsten-krypton
bulb.Thecameraacquirescolorinfraredimages(CIR)witharesolutionof1040x1392x3pixels
thatincludesNIRspectrumaswellasvisiblelightspectrum.Figure3.1Imagesequencecapturestartsatt0=0.Afterapproximately10seconds,att1,thelightsourceisturnedonilluminatingtheeyefor10moreseconds[t1;t2].Att2thelightsourceisturnedoffandremainsofffor10moreseconds[t2;t3].Thevideocaptureisstoppedatt3.3.3Description
Thedataiscollectedfrom54subjects,onevideo/eyewithanaverageof130frames/video.The
totalnumberofimagesis7115forthelefteyeand6985fortherighteyewithanaverageof440
pixelsacrosstheirisdiameter.ExampleofNIRimagesareshowninFigure3.2.DistributionofdemographicsandeyecolorinformationispresentedinTables3.1and3.2.Relationbetweenpupilradius(RP)andirisradius(RI)mayberepresentedasadifference,D,oraratio,R;whereD=RIRP;40Figure3.2SampleimagesfromthedatasetTable3.1DemographicsdistributionDemographicsCaucasian32Asian20African1AfricanAmerican1Table3.2EyecolorinformationEyeColorBlue7Green/Hazel6LightBrown/Mixed4Brown10DarkBrown27R=RPRI:R,isusuallyknownintheliteratureaspupildilationratio.Theirisradiusdoesnotchangeforalltheeyesevenwhenthepupilisundergoingdilationandconstriction.Hence,onlythepupilsizeis
foundtovarywhenthelightsourceisturnedonoroff.Figure3.3showsthehistogramofpupil
dilationratioofasubseton2218imagescorrespondingtothelefteyeinthedataset.410.20.250.30.350.40.450.50.550.60.65050100150200Pupil dilation ratio (R)CountFigure3.3Distributionofpupildilationratiosinthedataset.Theyrangefrom0.2177to0.6367.3.4Impactofpupildilation
PupildilationisknowntoimpactirismatchingsystemsbyincreasingtheHammingdistancebe-
tweenimagesofthesameeyehavingdifferentpupilsizes.Genuinescoresarecomputedforimages
ofthesameeyeatdifferentpupilsizesinordertostudytheimpactofpupildilation.Relationbe-
tweenthepupilandirisradiusforimagesI1andI2,denotedas(D1;R1)and(D2;R2),respectively,canbecomputedasfollows:D1=RI1RP1;D2=RI2RP2;R1=RP1RI1;R2=RP2RI2:Figure3.4showsthedistributionofgenuineHammingdistancescoresasafunctionof(a)jD1D2jand(b)jR1R2j.Typicalirisradiusisaround6mm.Thedifferenceiniriswidthsanddilationratiosarescaledwithrespectto6mmirisradiusandthreedifferentcategoriesof42dilationdifferences/ratiosareconsidered.Theboundariesbetweenthesecategoriescorrespondto
approximately0:5mm,1mmand>1mmdeformationinpupilradius.00.20.40.60.80500100015002000Difference of iris widths in pixelsCount  DD [0 22) pixelsDD [22 44) pixelsDD > 44 pixels)00.20.40.60.80500100015002000Difference of pupil dilation ratioCount  DR [0,0.0833]DR (0.0833,0.1667]DR > 0.1667(a)(b)Figure3.4DistributionofgenuineHammingdistancescoresasafunctionofdilationdifferences.
(a)jD1D2jand(b)jR1R2jItcanbeobservedfromalltheplotsinFigure3.4that,ingeneral,largerdifferencesiniriswidthsorpupildilationratiosresultinalargerHammingdistancewhenmatchingirisimagesof
thesameeye.Thissubstantiatesthepreviousofpupildilation'sadverseimpactoniris
matchingsystems.43CHAPTER4PROPOSEDMETHODSTheproposedmethodsrequireiristobeencodedusingdifferentofvaryingbandwidths.In
thiswork,unwrappedirisregionsareencodedusingmultiresolutionGaborThissection
describestheencodingprocesstogenerateIrisCode;themethodologyusedbytypicalirismatchers
togeneratematchscores;followedbytheproposednovelmatchingmethodandhowitisdifferent
fromthetypicalmatcher.
4.1Multi-resolutionGaborencoding
IrisCodescanbegeneratedbyapplyingmulti-scaleonanormalizedirisimageandquan-
tizingtheircomplexoutput.OnesuchimplementationbyOSIRISappliesofthreedifferent
sizes.EachproducestwobitsofIrisCodeperpixel.LettheithimagebedenotedbyIi.ItsnormalizedimageisdenotedasNi.Thesizeofthenormalizedimageisrtwhereristheradialresolutionandtistheangularresolution.ThreerectangularcomplexF1m1n1,F2m2n2andF3m3n3areappliedonthenormalizedimage.TheresultingcomplexoutputisthenconvertedtoabinaryIrisCodeset(C1i;C2i;C3i)r2talongwithamaskMir2t.Figure4.1pictoriallyshowsanIrisCodeset.Normalizedimagewithsizer=64andt=512forsizes915,927and951areusedinthiswork.Figure4.2showsanormalizedirisimageanditscorrespondingIrisCode
generatedusingthe3complexThesmallestencodessmallerregionsintheimage
andthelargestencodeslargerregionsintheimage.Thisisinthesmoothnessof
IrisCodesatdifferentsizes.ThelargerresultsinasmootherIrisCodecomparedtothe
smaller.44Normalized image  Code Code  Code  M IrisCode set Multi-scale filter encoding IrisCodes Mask Figure4.1Anormalizedimageisencodedusingmulti-scaletoresultinanIrisCodeset
alongwithamaskshowingvalidbitsineachIrisCode.Thismaskissameforallthecodesinthe
IrisCodesetNormalized Image Filter 1 
Filter 2 
Filter 3 Real Part Imaginary Part   Figure4.2AnormalizedimageanditscorrespondingIrisCodegeneratedusing3These
encodetheimageatmultiplescales.
4.2TypicalIrisCodematcher
LetussupposethatIrisCodesetsgeneratedfromtwonormalizedimagesNiandNjarebeingmatched.ThecorrespondingIrisCodesetsarerepresentedby(C1i;C2i;C3i;Mi)and(C1j;C2j;C3j;Mj)respectively.Acommonmask,MijiscomputedtodenotethelocationofcommonvalidbitscorrespondingtotheirisinboththeIrisCodes.45Mij=Mi\Mj:LettheresultofXORoperator,N,formatchingindividualIrisCodesgeneratedbyFbeRf:Rfij=CfiOCfj;f=1;2;3:Nresultsin0ifthecorrespondingbitsarethesameand1iftheyarenot.HammingdistancebetweentwoIrisCodesatthefthscaleisthengivenbyHDfij=kRfijTMijkkMijk;f=1;2;3:Typically,theHammingdistancescomputedforeacharefusedusingsumruletoproduceamatchingscore.Dijsum=HD1
ij+HD2
ij+HD3
ij:TheabovedescribedstepsemployedbyatypicalirismatcherarepresentedintheformofawchartinFigure4.3.
4.3Histogramofmatchingpatterns
Basedontheaforementioneddiscussion,threeoutputsareavailableateachpixellocation
inanirisimage.Hencethreematchingresults(r1;r2;r3)aregeneratedateverylocationwhentwoIrisCodesetsarematched.Thesethreeresultsateachlocationmaybecombinedand
representedasasinglevector,R,whichisreferredtoasmatchingbitpatternateverybitlocation.Itcanhavevaluessuchas000,001,010,...,111.Here,000atalocationwouldmeanthat
thepixelismatchedbyallscales;100wouldmeanthatalthoughthepixelismis-matched
at1,itismatchedby2and3.Similarly,111wouldindicatethatthepixelis
mis-matchedatallscales.46Code  Code  Result  M  Code  Code  Result  M  Code  Code  Result  M   Figure4.3Atypicalirismatcher.Matchscoresarecomputedindependentlyateachscalewhich
arethenfusedatscoreleveltoresultinadistancescore.Figure4.4showsdistributionofthesematchingpatternsforonesubject.Thelegendintheplotsdenotesthesizeofthepupilradiusinpixelsofthetwoimagesthatarebeingmatched.Itis
observedthatthepercentageof000s(matchedatalldecreaseswithincreaseindifference
ofpupildilationratiosbetweenthematchedsamples.Figure4.5showsdistributionsof
matchingpatternsforafewrandomlyselectedinter-class(impostor)pairsinthedataset.Itis
observedthatthedistributionofthesedecisionsisroughlyequalandsimilaracrossthedecision
patterns.Inatraditionalsumrulematcher,theinstancesof000,001,...,111wouldhavebeenmerelysummedupanddividedbythetotalnumberoflocations.Thiswouldmasksomeoftheinteresting
propertiesobservedinthesepatterns.Figure4.6showsthedistributionofthesematchingresults47Mean = 0.2085 Standard deviation = 0.0845 Peak at 000  Matched at all scales Figure4.4Distributionofdecisionsforgenuinematchingcasesforasinglesubject.atthreedifferentscalesforthegenuineandimpostorcases.ItisobservedfromFigure4.6thatsomematchingpatterns,suchas000,011,101,110and111,aremuchmorediscriminativecomparedtoothers.Hence,thesedecisionscouldbe
selectivelyfusedtoprovidebetterperformance.
4.4Fusion
Theideabehindtheproposedmethodistomakeamatchingdecisionateachpixellocationbasedoninformationatmultiplescales.Thedistributionofdecisionpatternsshownintheprevious
sectionsareexploitedtocomeupwithabetterdecisionstrategy.IrisCodebitsgeneratedfrom
multipleareselectivelymatchedtocomputealdissimilarityscore.Thisispictorially48 Mean = 0.4858 Standard deviation = 0.0446  
 Uniformly distributed Figure4.5DistributionofdecisionsforrandomlyselectedimpostormatchingcasesdepictedinFigure4.7.
4.4.1RulebasedFusion
Multipledecisionstrategiescanbedevelopedtoallowforstrcitorrelaxedmatchingconditions.
Theproposedmatchingstrategiesaredescribedbelow.
Method1:Twoirisimages(Ii;Ij)arematchedusingIrisCodesgeneratedby1ateachbitlo-cation,r1=(c1
i;c1j).Iftheimagesarenotmatchedat1,i.e,r1=1,thenthematchingisextendedtoIrisCodesgeneratedbylarger2and3.Thebitlocationisdeemeda
match,ifIrisCodesareatleastmatchedbyters2and3.Thishelpsinhandlinglocaldefor-49Figure4.6Comparisonofdistributionsofpossibledecisionsforgenuineandimpostor
matchingcasesmationssincematchisestablishedatalargerscaleforthosebitsthatwouldhaveotherwisemismatchedatsmallerscales.Method2:Thismethodrelaxestheconditionsforamatch.IftwoIrisCodesarenotmatchedatthe
lowestscale,anadditionalopportunityisprovidedatmediumscale2.IncaseIrisCodes
arenotmatchedat2,thenaopportunityisaffordedatlarger3.Thismethod50Code  Code  Result  Code  Code  Result  Code  Code  Result  Sequential fusion Result  M  Figure4.7Theproposedirismatchersequentiallycombinestheresultsatmultiplescalesand
generatesasingledecisionresult.allowsforapositivematchiftheirisregionsarematchedatleastinoneofthescales.Method3:Thismethodprovidesastrictermatchingcriterioncomparedtoalltheothermethodsby
requiringtheIrisCodestomatchat1aswellaseither2or3.Thismethod
removesthepossibilityofmatchinglocallydeformedregions.Onlythoseregionsthatarematchedatmultiplescalesaredeemedamatch.ThelogicaloperationsshowninFigures4.8,4.9and4.10areusedinthesequentialfusionstepinFigure4.7andcanbeimplementedusingasingleBooleanexpression.Correspondingtruth
tablesareusedtoderivetheBooleanexpressionthatdirectlycomputestheresultbasedonthe51Table4.1LogicaloperationsusedtocombinetheoutputofmultipleIrisCodes.FusionLogicSumruleR1+R2+R3Method1R1&(R2j(˘R2&R3))Method2R1&R2&R3Method3R1j(˘R1&R2&R3)decisionsateachscale.Hence,asingledecisionismade,r=0(match)andr=1(non-match),ateachbitlocationinanIrisCode.Thedecisionisequivalenttoapplyingasinglecomplex
onthenormalizedimage.LetthematchingdecisionbitsbepresentedinamatrixR.HammingDistancebetweentwoIrisCodesets(C1i;C2i;C3i;Mi)and(C1j;C2j;C3j;Mj)isthengivenbyDij=krijTMijkkMijk:Table4.1showsthelogicaloperationsforthesethreemethodsalongwiththesimplesumrule
fusion.
4.4.2basedFusion
Asseenintheprevioussection,ahistogramofmatchingpatternsisbeinggeneratedforevery
pairofimagesthatarebeingmatched.AlinearSVMwastrainedusinghistogramsof
matchingpatternsforgenuineandimpostorcasesonatrainingdataset.Givenanewpairofiris
images,thetrainedwasusedtopredictifthenewhistogramofmatchingpatternspertains
tothegenuineoraimpostorcase.TheobtainedresultswerefoundtocomparabletotheMethod1
proposedintheprevioussection.However,furtherresearchonthistopicwillbenecessary.
4.5ExperimentsandResults
Theproposedmethodsaretestedonlefteyeimagesacquiredatfullilluminationintheproprietary
pupildilationdataset.Atotalof2218imagesoflefteyesfrom52subjectsisusedtotestthe
proposedmethods.Theimagesareautomaticallysegmented,normalizedandencodedusingthe52Filter 3 bit Match   Filter 2 bit Match Filter 1 bit Match Match Non-match Method 1     0 0 0 0 
0 0 1 0 
0 1 0 0 
0 1 1 0 
1 0 0 0 
1 0 1 1 
1 1 0 1 
1 1 1 1  Yes No  Yes Yes No  No  Method 1 applied  at each bit location Figure4.8FlowchartdepictingMethod1anditscorrespondingtruthtableOSIRIS_v4.1SDK.SemilogROCsarepresentedtobetterobservetheperformanceatlowFARs.
Atotalof46,480genuinescoresand1,696,504impostorscoresaregenerated.Figure4.11(a)
showsROCsforthefulldata.Itisclearlyseenthatallthethreemethodsclearlyimproveupon
thetraditionalsumrulefusionmethod.However,genericmatchingusingMasek's1-Dencoded
IrisCodes[4]isobservedtoprovidebetterstandaloneperformance.Judiciousparametertun-
ingusing2-DGaborwouldprobablyyieldbetterperformance,inwhichcasetheproposed
methodisexpectedtofurtherimprovetheperformance.Itcanalsobeobservedthatfusingscores
fromMethod1withmatchscoresfromMasek's1-DencodedIrisCoderesultsintheoverallbest
performance.Inordertoobservetheimpactoftheproposedmethodsondeformedirispatterns,scoresfromthetraditionalmatchingmethodsandproposedmethodsbasedondifferencesinpupildilation53Filter 3 bit Match Filter 2 bit Match Filter 1 bit Match   Match Non-match Method 2     0 0 0 0 
0 0 1 0 
0 1 0 0 
0 1 1 0 
1 0 0 0 
1 0 1 0 
1 1 0 0 
1 1 1 1  Yes No  Yes Yes No  No  Method 2 applied  at each bit location Figure4.9FlowchartdepictingMethod2anditscorrespondingtruthtableratioareexamined.Thegenuinescoresaredividedintothreedilationgroups-small,mediumand
large-dependingontheabsolutevalueofthedifferenceinpupildilationratiobetweenthepairof
imagesbeingmatched.Impostordistributionsarekeptthesamefortherespectivemethods.These
ROCsareshowninFigure4.12.ItisevidentfromtheROCplotsinFigure4.12thattheproposed
methodshavealargerimpactwhencomparinghighlydeformedpatternsthanwhencomparing
twoimageswithalmostthesamepupildilationvalues.FusingbestperformingMethod1with
Masek1-Dmethod[4]resultsinthebestoverallperformancewhencomparingimageswithlarger
differencesinpupilsizes.Figure4.13showsthehistogramdistributionsofgenuineandimpostor
scoresforMasek'smethodaloneandafterfusingtheMasek'sscorewiththematchscorefrom
Method1.Thesematchingmethodsarenotjustlimitedtohandlingdeformationduetopupildilation/constriction54Method 3     0 0 0 0 
0 0 1 0 
0 1 0 0 
0 1 1 1 
1 0 0 1 
1 0 1 1 
1 1 0 1 
1 1 1 1  Method 3 applied  at each bit location Filter 3 bit Match Yes Filter 2 bit Match Filter 1 bit Match   Match Non-match No  Yes Yes No  No  Figure4.10FlowchartdepictingMethod3anditscorrespondingtruthtablealone,butcanbeusedtohandlenon-idealirisimages.Tovalidatetheefyofthesemethods,
experimentswereconductedontheWVUnonideal[82]andQFire[83]datasetsaswell.The
WVUnon-idealdatasethas1557imagesfrom241subjectsobtainedundernon-idealconditions
exhibitingthepresenceofblur,outoffocusandocclusion.Atotalof5277genuinescoresand
1206069impostorscoresaregeneratedontheWVUdataset.QFirehas1304lefteyeimagesfrom
90subjectsimagedatvariousacquisitiondistances.Atotalof8847genuinescoresand840709
impostorscoresaregeneratedontheQFiredataset.Figure4.14showstheresultofapplyingthe
proposedmatchingmethodsonWVUandQFiredatasets,andtheimprovementinperformanceis
clearlyobserved.5510-410-21001029596979899100False Accept Rate (%)Genuine Accept Rate (%)  Sum ruleMasek-1DOSIRIS-SDKMethod1Method2Method3Method 1 + Masek-1D00.10.20.30.40.50.60.700.020.040.060.080.10.120.14Hamming DistanceNormalized histograms of IrisCode match scores for Method 1  Genuine Sum ruleImpostor Sum ruleGenuine Method1Impostor Method1Decrease indistance scoreDecrease indistance score(a)(b)-0.100.10.20.30.40.50.600.050.10.150.20.250.30.35Hamming DistanceNormalized histograms of IrisCode match scores for Method 2  Genuine Sum ruleImpostor Sum ruleGenuine Method2Impostor Method200.10.20.30.40.50.60.700.020.040.060.080.10.120.14Hamming DistanceNormalized histograms of IrisCode match scores for Method 3  Genuine Sum ruleImpostor Sum ruleGenuine Method3Impostor Method3Increase indistance scoreIncrease indistance score(c)(d)Figure4.11(a)ROCsforfulldata.Thegenuineandimpostorscoredistributionsareplottedfor
(b)Method1,(c)Method2and(d)Method3.
4.6Examples
Theproposedrule-basedmatchingmethodisabletoprovidebettervperformanceover
thetraditionalmethod.ThisimpliesthatatalowoperatingFARof,say,0:0001%FARadilatedprobeimagethatwouldnothavepreviouslymatchedwithanon-dilatedimageinthegallerywould
nowbecorrectlyidentiusingthenewmatchingscheme.Examplesofsuchpairsofimagesare
showninFigure4.15.Itishoweverpossiblethattheimprovementmaynotbeapparent,orcan
resultinafalsemis-matchwhencomparinggenuinepairimageswithsimilarpupilsizebutlarge
Hammingdistance(duetoocclusion/specular5610-410-21001029596979899100False Accept Rate (%)Genuine Accept Rate (%)Small difference in iris width  Sum RuleMethod1Method2Method3Masek-1DMethod1 + Masek-1D10-410-210010280859095100False Accept Rate (%)Genuine Accept Rate (%)Medium difference in iris width  Sum RuleMethod1Method2Method3Masek-1DMethod1 + Masek-1D(a)(b)10-410-2100102707580859095100False Accept Rate (%)Genuine Accept Rate (%)Large difference in iris width  Sum RuleMethod1Method2Method3Masek-1DMethod1 + Masek-1D(c)Figure4.12ROCsgeneratedbyusingthegenuinescoresforpairswhosepupildilationratio
differencesare(a)small,(b)mediumand(c)large.Theimpostordistributionsareheldthesame
acrossallthecases.57Figure4.13ThehistogramofgenuineandimpostorscoresusingMasek'smethodandafter
fusionofmatchscoresfromMasek'smethodandproposedMethod1.10-310-210-1100101102020406080100False Accept Rate (%)Genuine Accept Rate (%)ROCs for WVU non-ideal dataset  OSIRIS Sum ruleMasek-1DOSIRIS-SDKMethod1Method2Method310-310-210-1100101102556065707580859095100False Accept Rate (%)Genuine Accept Rate (%)ROCs for QFire dataset  OSIRIS sum ruleMasek-1DOSIRIS-SDKMethod1Method2Method3(a)(b)Figure4.14ROCcurvesfor(a)WVUand(b)QFiredatasets.TheimprovementinGARisclearly
evidentatlowFARs.58ProbeImageGalleryImageFigure4.15Genuinepairsofimagesthatwerecorrectlymatchedusingtheproposedmethodbut
wereincorrectlyrejectedbythetraditionalmatchingmethodat0:0001%FAR.59CHAPTER5SUMMARY5.1Summary
AnovelselectivematchingschemebasedonIrisCodesobtainedusingmultiscaleispro-
posed.Thethreeproposedmethodsintegratethedecisionsmadeatthebi(orpixel)level,thereby
accountingforlocaldeformations.Theproposedapproachisshowntoimprovethe
matchingaccuracywhencomparingimageswithlargedifferencesinpupildilationratio.Itisalso
showntoimprovetheperformancewhennon-idealiridesarecomparedwheremultiplefactors
includingimpropersegmentation,off-gazeimagescouldnegativelymatchingaccuracy.
Futureworkwillaimatexploringothermatchingstrategies,whicharebasedonadeeperunder-
standingoftheadvantageofthesemethods.Thedistributionofmultidecisionpatternscouldbeusedasafeaturevectorandaclascouldbetrainedtoselectthebestdecisionstrategy.Inthiswork,sizesareincreasedalongthe
angulardirection;infuture,weaimtoexploreothersetssuchasthosevaryingincreasingin
theradialdirection,orthoseradialandangularetc.Inaddition,thepossibilityofdesigning
newencodingschemesbasedontheresultsofthisthesiswillbeexplored.Inparticular,bit-level
decisionintegrationcanallowforthegenerationofspatialstatisticsthatcanpotentiallyprovide
deeperinsightsintothevariationoftheiristextureofanindividual.60BIBLIOGRAPHY61BIBLIOGRAPHY[1]HelgaKolb,fiWebvision,flhttp://www.nextgenid.com/.[2]KebesatEnglishWikipedia,fiAbsorptionspectrumofwater,flhttp://commons.wikimedia.org/wiki/File:Absorption_spectrum_of_liquid_water.png.[3]NikiforosKollias,fiThespectroscopyofhumanmelaninpigmentation,flinJournalofInves-tigativeDermatology.BLACKWELLSCIENCEINC238MAINST,CAMBRIDGE,MA02142,1994,vol.102,pp.268Œ268.[4]L.Masek,fiMATLABsourcecodeforaBiometricSystemBasedonIrisPat-terns,flTheSchoolofComputerScienceandSoftwareEngineering,TheUniversityofWesternAustralia,2003.[5]Z.Wei,T.Tan,andZ.Sun,fiNonlinearirisdeformationcorrectionbasedongaussianmodel,flAdvancesinBiometrics,pp.780Œ789,2007.[6]K.Hollingsworth,K.W.Bowyer,andP.J.Flynn,fiPupildilationdegradesirisbiometricperformance,flComputerVisionandImageUnderstanding,vol.113,no.1,pp.150Œ157,2009.[7]H.Rohen,fiDerbauderregenbogenhautbeimmenschenundeinigensäugern,flGegenbaurMorphologyJournal,vol.91,pp.140,1951.[8]H.J.Wyatt,fiA`minimum-wear-and-tear'meshworkfortheiris,flVisionResearch,vol.40,no.16,pp.2167Œ2176,2000.[9]X.YuanandP.Shi,fiAnon-linearnormalizationmodelforirisrecognition,flAdvancesinBiometricPersonAuthentication,pp.135Œ141,2005.[10]AgeevatWikimediaCommons,fiImageofhumaneye,flhttp://commons.wikimedia.org/wiki/File:Ageev_iris.jpg.[11]AnilKJain,KarthikNandakumar,andArunRoss,fi50yearsofbiometricresearch:Accom-plishments,challenges,andopportunities,flPatternRecognitionLetters,2016.[12]fiISO/IEC2382-37:2012,flhttp://www.iso.org/iso/home/store/catalogue_tc/catalogue_detail.htm?csnumber=55194.[13]DineiFlorencioandCormacHerley,fiAlarge-scalestudyofwebpasswordhabits,flinPro-ceedingsofthe16thinternationalconferenceonWorldWideWeb.ACM,2007,pp.657Œ666.[14]A.K.Jain,A.Ross,andS.Prabhakar,fiAnintroductiontobiometricrecognition,flIEEETransactionsonCircuitsandSystemsforVideoTechnology,vol.14,no.1,pp.4Œ20,2004.[15]BrianDeCannandArunRoss,fifihasthispersonbeenencounteredbefore?fl:Modelingananonymoussystem,flinComputerVisionandPatternRecognitionWorkshops(CVPRW).IEEE,2012,pp.89Œ96.62[16]NathanDKalka,JinyuZuo,NataliaASchmid,andBojanCukic,fiImagequalityassessmentforirisbiometric,flinDefenseandSecuritySymposium.InternationalSocietyforOpticsandPhotonics,2006,pp.62020DŒ62020D.[17]J.Daugman,fiHowirisrecognitionworks,flIEEETransactionsonCircuitsandSystemsforVideoTechnology,vol.14,no.1,pp.21Œ30,2004.[18]RezaDerakhshani,ArunRoss,andSimonaCrihalmeanu,fiAnewbiometricmodalitybasedonconjunctivalvasculature,flProc.ofNeuralNetworksinEngineering(ANNIE),pp.1Œ8,2006.[19]RobertBHillandJoelCJohnson,fiEyefundusopticalscannersystemandmethod,flJuly21996,USPatent5,532,771.[20]PawelKasprowskiandJózefOber,fiEyemovementsinbiometrics,flinBiometricAuthenti-cation,pp.248Œ258.Springer,2004.[21]CoreyDHollandandOlegVKomogortsev,fiBiometricvviacomplexeyemove-ments:Theeffectsofenvironmentandstimulus,flinBiometrics:Theory,ApplicationsandSystems(BTAS).IEEE,2012,pp.39Œ46.[22]UnsangPark,ArunRoss,andAnilKJain,fiPeriocularbiometricsinthevisiblespectrum:Afeasibilitystudy,flinBiometrics:Theory,Applications,andSystems,2009.IEEE,2009,pp.1Œ6.[23]RichardASturmandMatsLarsson,fiGeneticsofhumaniriscolourandpatterns,flPigmentcell&melanomaresearch,vol.22,no.5,pp.544Œ562,2009.[24]SunpreetSArora,MayankVatsa,RichaSingh,andAnilJain,fiIrisrecognitionunderalcoholApreliminarystudy,flin5thIAPRInternationalConferenceonBiometrics(ICB).IEEE,2012,pp.336Œ341.[25]V.F.Pamplona,M.M.Oliveira,andG.V.G.Baranoski,fiPhotorealisticmodelsforpupillightxandiridalpatterndeformation,flACMTransactionsonGraphics(TOG),vol.28,no.4,pp.106,2009.[26]HugoProença,fiOnthefeasibilityofthevisiblewavelength,at-a-distanceandon-the-moveirisrecognition,flinComputationalIntelligenceinBiometrics:Theory,Algorithms,andAp-plications.IEEE,2009,pp.9Œ15.[27]C.Boyce,A.Ross,M.Monaco,L.Hornak,andX.Li,fiMultispectralirisanalysis:Apre-liminarystudy,flinCVPRworkshoponBiometrics,2006.[28]A.Ross,R.Pasula,andL.Hornak,fiExploringmultispectralIrisrecognitionbeyond900nm,flinInternationalConferenceonBiometrics:Theory,Applications,andSystems(BTAS).IEEE,2009,pp.1Œ8.63[29]FrederickWWheeler,AGAmithaPerera,GilAbramovich,BingYu,andPeterHTu,fiStand-offirisrecognitionsystem,flinBiometrics:Theory,ApplicationsandSystems.IEEE,2008,pp.1Œ7.[30]WenboDong,ZhenanSun,andTieniuTan,fiAdesignofirisrecognitionsystematadistance,flinChineseConferenceonPatternRecognition.IEEE,2009,pp.1Œ5.[31]SRIinternational,fiWeblink,flhttp://www.sri.com/engage/products-solutions/iom-passport-portal-system.[32]SRIinternational,fiWeblink,flhttp://www.sri.com/engage/products-solutions/iom-pass-thru-drive-up-iris-recognition-system.[33]VishnuNareshBoddetiandBVKVijayaKumar,irisrecognitionusingunrestoredwavefront-codedimagery,flIEEETransactionsonSystems,ManandCy-bernetics,PartA:SystemsandHumans,,vol.40,no.3,pp.495Œ508,2010.[34]SriramPTankasala,VikasGottemukkula,SashiKanthSaripalle,VenkataGoutamNalamati,RezaDerakhshani,RaghunandanPasula,andArunRoss,fiAvideo-basedhyper-focalimag-
ingmethodforirisrecognitioninthevisiblespectrum,flinIEEEConferenceonTechnologiesforHomelandSecurity(HST),2012,pp.214Œ219.[35]RichardPWildes,fiIrisrecognition:anemergingbiometrictechnology,flProceedingsoftheIEEE,vol.85,no.9,pp.1348Œ1363,1997.[36]L.Ma,Y.Wang,andT.Tan,fiIrisrecognitionusingcircularsymmetricflinProceed-ingsofthe16thInternationalConferenceonPatternRecognition(ICPR),2002,vol.2,pp.414Œ417.[37]JinyuZuoandNataliaASchmid,fiOnamethodologyforrobustsegmentationofnonidealirisimages,flIEEETransactionsonSystems,Man,andCybernetics,,vol.40,no.3,pp.703Œ718,2010.[38]A.RossandS.Shah,fiSegmentingnon-idealirisesusinggeodesicactivecontours,flinBio-metricsSymposium:SpecialSessiononResearchattheBiometricConsortiumConference,2006,pp.1Œ6.[39]KaushikRoy,PrabirBhattacharya,andChingYSuen,fiUnidealirissegmentationusingregion-basedactivecontourmodel,flinImageAnalysisandRecognition,pp.256Œ265.Springer,2010.[40]MayankVatsa,RichaSingh,andAfzelNoore,fiImprovingirisrecognitionperformanceusingsegmentation,qualityenhancement,matchscorefusion,andindexing,flIEEETransactionsonSystems,Man,andCybernetics,PartB:Cybernetics,vol.38,no.4,pp.1021Œ1035,2008.[41]RandyPBroussard,LaurenRKennell,DavidLSoldan,andRobertWIves,fiUsingneuralnetworksandfeaturesaliencytechniquesforimprovedirissegmentation,flinNeuralNetworks,2007.IJCNN2007.InternationalJointConferenceon.IEEE,2007,pp.1283Œ1288.64[42]ZhaofengHe,TieniuTan,ZhenanSun,andXianchaoQiu,fiTowardaccurateandfastirissegmentationforirisbiometrics,flIEEETransactionsonPatternAnalysisandMachineIntel-ligence,vol.31,no.9,pp.1670Œ1684,2009.[43]RuihuiZhu,JinfengYang,andRenbiaoWu,fiIrisrecognitionbasedonlocalfeaturepointmatching,flinInternationalSymposiumonCommunicationsandInformationTechnologies.IEEE,2006,pp.451Œ454.[44]FernandoAlonso-Fernandez,PedroTome-Gonzalez,VirginiaRuiz-Albacete,andJavierOrtega-Garcia,fiIrisrecognitionbasedonsiftfeatures,flinInternationalConferenceonBiometrics,IdentityandSecurity(BIdS).IEEE,2009,pp.1Œ8.[45]WageehWBolesandBoualemBoashash,fiAhumantechniqueusingimagesoftheirisandwavelettransform,flIEEEtransactionsonsignalprocessing,vol.46,no.4,pp.1185Œ1188,1998.[46]Chia-TeChou,Sheng-WenShih,Wen-ShiungChen,andVictorWCheng,fiIrisrecognitionwithmulti-scaleedge-typematching,flin18thInternationalConferenceonPatternRecogni-tion.IEEE,2006,vol.4,pp.545Œ548.[47]DdeMartin-Roche,CarmenSanchez-Avila,andRaulSanchez-Reillo,fiIrisrecognitionforbiometricusingdyadicwavelettransformzero-crossing,flin35thInternationalCarnahanConferenceonSecurityTechnology.IEEE,2001,pp.272Œ277.[48]VivekanandDorairaj,NataliaASchmid,andGamalFahmy,fiPerformanceevaluationofiris-basedrecognitionsystemimplementingpcaandicaencodingtechniques,flinDefenseandSecurity.InternationalSocietyforOpticsandPhotonics,2005,pp.51Œ58.[49]Ya-PingHuang,Si-WeiLuo,andEn-YiChen,fiAnefirisrecognitionsystem,flinInternationalConferenceonMachineLearningandCybernetics.IEEE,2002,vol.1,pp.450Œ454.[50]LiMa,TieniuTan,YunhongWang,andDexinZhang,fiPersonalbasedoniristextureanalysis,flIEEETransactionsonPatternAnalysisandMachineIntelligence,vol.25,no.12,pp.1519Œ1533,2003.[51]Wen-ShiungChen,Ren-HungHuang,andLiliHsieh,fiIrisrecognitionusing3dco-occurrencematrix,flinAdvancesinBiometrics,pp.1122Œ1131.Springer,2009.[52]MelvinEmber,EncyclopediaofSexandGender:MenandWomenintheWorld'sCulturesTopicsandCulturesAK-Volume1;CulturesLZ,vol.2,Springer,2003.[53]K.W.Bowyer,K.Hollingsworth,andP.J.Flynn,fiImageunderstandingforirisbiometrics:Asurvey,flComputerVisionandImageUnderstanding,vol.110,no.2,pp.281Œ307,2008.[54]KevinWBowyer,SarahEBaker,AmandaHentz,KarenHollingsworth,TanyaPeters,andPatrickJFlynn,fiFactorsthatdegradethematchdistributioninirisbiometrics,flIdentityintheInformationSociety,vol.2,no.3,pp.327Œ343,2009.65[55]RegisteredTravelerInteroperabilityConsortium(RTIC),fiTechnicalinteroperabilitycationfortheRegisteredTravelerProgram(Version1.0ŒFinal),flOnline,2006.[56]SoumyadipRakshitandDonaldMMonro,fiAnevaluationofimagesamplingandcompres-sionforhumanirisrecognition,flIEEETransactionsonInformationForensicsandSecurity,vol.2,no.3,pp.605Œ612,2007.[57]JohnDaugmanandCathrynDowning,fiEffectofsevereimagecompressiononirisrecogni-tionperformance,flIEEETransactionsonInformationForensicsandSecurity,vol.3,no.1,pp.52Œ61,2008.[58]IshanNigam,MayankVatsa,andRichaSingh,fiOcularbiometrics:Asurveyofmodalitiesandfusionapproaches,flInformationFusion,vol.26,pp.1Œ35,2015.[59]GeorgeMcConnon,FarzinDeravi,SanaulHoque,KonstantinosSirlantzis,andGarethHow-ells,fiImpactofcommonophthalmicdisordersonirissegmentation,flin5thIAPRInterna-tionalConferenceonBiometrics(ICB).IEEE,2012,pp.277Œ282.[60]MassimoTistarelliandMarkSNixon,AdvancesinBiometrics:ThirdInternationalConfer-ences,ICB2009,Alghero,Italy,June2-5,2009,Proceedings,vol.5558,Springer,2009.[61]DésiréeWhiteandMontserratRabago-Smith,fiGenotypeŒphenotypeassociationsandhumaneyecolor,flJournalofhumangenetics,vol.56,no.1,pp.5Œ7,2011.[62]SamuelPFenkerandKevinWBowyer,fiExperimentalevidenceofatemplateagingeffectinirisbiometrics,flinWorkshoponApplicationsofComputerVision(WACV).IEEE,2011,pp.232Œ239.[63]IreneELoewenfeldandOttoLowenstein,Thepupil:anatomy,physiology,andclinicalapplications,vol.2,Wiley-Blackwell,1993.[64]G.Francois,P.Gautron,G.Breton,K.Bouatouch,etal.,fiAnatomicallyaccuratemodelingandrenderingofthehumaneye,fl2007.[65]G.Francois,S.Pattanaik,K.Bouatouch,andG.Breton,fiSubsurfacetexturemapping,flSIGGRAPH,2006.[66]AntwanDClark,ScottAKulp,IsomHHerron,andAARoss,fiAtheoreticalmodelfordescribingirisdynamics,flinHandbookofIrisRecognition,pp.129Œ150.Springer,2013.[67]A.N.PalazottoandS.T.Dennis,Nonlinearanalysisofshellstructures,AIAAEducationServices,1992.[68]W.F.Carroll,APrimerforFiniteElementsinElasticStructures,JohnWiley&Sons,Inc,1999.[69]RSGejji,ADClark,SCrihalmeanu,andAARossy,fiUnderstandingtheeffectsofpupildilationonirisrecognitioninthenirspectrum,flinIEEEInternationalSym-posiumonTechnologiesforHomelandSecurity(HST).IEEE,2015,pp.1Œ6.66[70]ADClark,RSGejji,ADRidley,andAARoss,fiThestatesofdilationinirisrecognition(apreliminarystudy),flinIEEEInternationalSymposiumonTechnologiesforHomelandSecurity(HST).IEEE,2015,pp.1Œ5.[71]D.TerzopoulosandD.Metaxas,fiDynamic3dmodelswithlocalandglobaldeformations:Deformablesuperquadrics,flinTransactionsonPatternAnalysisandMachineIntelligence.IEEE,1990,vol.13.[72]M.Mooney,fiAtheoryoflargeelasticdeformation,flJournalofAppliedPhysics,vol.11,no.9,pp.582Œ592,1940.[73]F.L.Bookstein,fiPrincipalwarps:Thin-platesplinesandthedecompositionofdeformations,flIEEETransactionsonPatternAnalysisandMachineIntelligence,,vol.11,no.6,pp.567Œ585,1989.[74]JasonThornton,MariosSavvides,andVKumar,fiAbayesianapproachtodeformedpatternmatchingofirisimages,flIEEETransactionsonPatternAnalysisandMachineIntelligence,vol.29,no.4,pp.596Œ606,2007.[75]InmaculadaTomeo-Reyes,ArunRoss,AntwanDClark,andVinodChandran,fiAbiome-chanicalapproachtoirisnormalization,flinInternationalConferenceonBiometrics(ICB).IEEE,2015,pp.9Œ16.[76]LDhir,NEHabib,DMMonro,andSRakshit,fiEffectofcataractsurgeryandpupildilationonirispatternrecognitionforpersonalauthentication,flEye,vol.24,no.6,pp.1006Œ1010,2009.[77]OrangSeyeddain,HannesKraker,AndreasRedlberger,AloisKDexl,GuentherGrabner,andMartinEmesz,fiReliabilityofautomaticbiometricirisrecognitionafter
ordrug-inducedpupildilation,flEuropeanjournalofophthalmology,vol.24,no.1,pp.58Œ62,2013.[78]KarenPHollingsworth,KevinWBowyer,andPatrickJFlynn,fiThebestbitsinaniriscode,flTransactionsonPatternAnalysisandMachineIntelligence,,vol.31,no.6,pp.964Œ973,2009.[79]ChristianRathgeb,AndreasUhl,andPeterWild,fiOncombiningselectivebestbitsofiris-codes,flinBiometricsandIDManagement,pp.227Œ237.Springer,2011.[80]ChristianRathgeb,AndreasUhl,andPeterWild,fiShiftingscorefusion:Onexploitingshiftingvariationinirisrecognition,flinProceedingsofthe2011ACMSymposiumonAppliedComputing.ACM,2011,pp.3Œ7.[81]JamesEGentile,NaliniRatha,andJonathanConnell,fiAneftwo-stageirisrecognitionsystem,flinBiometrics:Theory,Applications,andSystems.IEEE,2009,pp.1Œ5.[82]S.Crihalmeanu,ArunRoss,StephanieSchuckers,andL.Hornak,fiAprotocolformultibio-metricdataacquisition,storageanddissemination,flWVU,LaneDept.ofComputerScienceandElectricalEng.,Morgantown,WV,TechicalReport,2007.67[83]S.Schuckers,P.Lopez,P.Johnson,N.Sazonova,F.Hua,R.Lazarick,C.Miles,E.Talbassi,E.Sazonov,A.Ross,andL.Hornak,fiQualityŒface/irisresearchensembledataset
overview,flClarksonUniversity,DeptofElectricalandComputerEngineering,TechicalReport,2010.68